In December 2024, Anthropic hit $1 billion in annualized revenue. A well-funded AI lab, sure. Interesting research, decent product, strong team. But a billion dollars in AI is table stakes. OpenAI was pulling in $13 billion. Google was spending more on AI infrastructure in a single quarter than Anthropic had earned in its entire existence.

Yesterday, Bloomberg reported that Anthropic’s annual run rate has passed $30 billion. That’s 30x growth in 16 months. SaaStr called it “simply without precedent in B2B software. Not Slack, not Zoom, not Snowflake.”

The Founding (2021)

In late 2020, eight people walked out of OpenAI.

Dario Amodei was VP of Research, the person who oversaw GPT-2 and GPT-3 and co-invented RLHF (Reinforcement learning from human feedback), the training technique behind essentially every AI assistant you use today. His sister Daniela was VP of Safety and Policy. But this wasn’t just the Amodei siblings. Tom Brown had led the engineering behind GPT-3. Jared Kaplan co-authored the scaling laws paper that mathematically proved bigger models get predictably better. Chris Olah was pioneering the field of mechanistic interpretability, the attempt to understand what’s actually happening inside neural networks. Jack Clark had been OpenAI’s policy director. Sam McCandlish co-led the scaling laws research with Kaplan. Ben Mann rounded out the group.

Dario has been explicit about why: “People say we left because we didn’t like the deal with Microsoft. False.” The group believed that pouring more compute into models would make them dramatically more capable, and that capability alone was dangerous without alignment research keeping pace. They decided OpenAI’s leadership didn’t share that second conviction with the same intensity.

They incorporated Anthropic in January 2021 as a Public Benefit Corporation. The name means “relating to human existence.” They picked it from a spreadsheet that also included “Sponge,” “Sloth,” and “Sparrow Systems.” I am quite glad they picked Anthropic.

Early Days and Constitutional AI (2021-2023)

The company was founded in the middle of COVID’s second wave. Everyone met on Zoom. Eventually the early team of fifteen or twenty people would meet for weekly lunches in San Francisco’s Precita Park, pulling up their own chairs to talk business.

Three months after incorporating, they raised $124 million in a Series A from Jaan Tallinn (Skype co-founder), Dustin Moskovitz (Facebook co-founder), and Eric Schmidt. The team immediately started working on Constitutional AI. Jared Kaplan’s description: “We’re just gonna write a constitution for a language model and that’ll change all of its behavior.” Give the AI a written set of principles. Have it critique and revise its own outputs. No army of human labelers needed. They published the paper in December 2022, and it introduced a technique (RLAIF) that has since become a default across the industry. But in those first months, they were a small team in a park with a thesis and some funding.

Then came the Series B in April 2022: $580 million, led by Sam Bankman-Fried. Alameda Research put in $500 million of the $580 million (yes, the infamous FTX). Soon after, SBF was convicted of fraud and sentenced to 25 years. The FTX estate eventually sold its Anthropic stake for about $1.4 billion in 2024.

Amazon invested $8 billion total across multiple tranches, making AWS Anthropic’s primary cloud partner. Google put in over $3 billion and now owns 14% of the company.

The Model Race (2023-2024)

Claude 1 launched in March 2023, the same month as GPT-4. Safer and more careful, but clearly less capable. Claude 3 in March 2024 changed things. Three models launched simultaneously, and Opus beat GPT-4 on multiple benchmarks. First time a non-OpenAI model topped the Chatbot Arena leaderboard. Revenue went from roughly $100 million at the start of 2024 to $1 billion by December.

But the real turn came with Claude 3.5 Sonnet in June 2024. A mid-tier model outperformed the flagship Opus at 80% lower cost. It scored 64% on agentic coding tasks versus Opus’s 38%. This was the moment developers stopped treating Claude as “the safety AI” and started treating it as the best coding AI. Cursor made it their default model. I can’t think of another example in software where a company’s second-tier product cannibalized its own flagship and everyone, including the company, was fine with it.

Claude Code: The Accidental Rocket Ship (2024-2025)

Claude code was an accident.

In September 2024, Boris Cherny joined Anthropic. He’d been a principal engineer at Meta. His first prototype was a trivial music control tool. Then he gave Claude access to the filesystem, and described discovering “product overhang”: the model already had the capability to be a genuine development partner, but no product existed to surface it. When Claude could read files, it didn’t just answer questions. It explored. It followed imports. It understood project structure without being told to.

Boris joined Anthropic with a team of three. The product he built now accounts for roughly a fifth of a $30 billion revenue machine.

Claude Code launched as a research preview in February 2025. Six months after going GA, it hit $1 billion in ARR. Slack took seven years to reach $1 billion. By February 2026, it had crossed $2.5 billion. SemiAnalysis reported that 4% of all public GitHub commits were being authored by Claude Code. And Microsoft, the company that sells GitHub Copilot, internally adopted Claude Code across major engineering teams. That’s the AI equivalent of catching Coca-Cola executives drinking Pepsi in the break room.

In a 15,000-developer survey, Claude Code earned a 46% “most loved” rating versus Cursor at 19% and Copilot at 9%. The tool now writes 90% of its own codebase. Boris’s team philosophy: “Every time there’s a new model release, we delete a bunch of code.”

The Enterprise Explosion (2025-2026)

85% of Anthropic’s revenue comes from enterprise customers. Claude has roughly 18.9 million monthly active users, about 5% of ChatGPT’s base, but it generates revenue equivalent to 40% of OpenAI’s. An 8x monetization advantage per user. Over 1,000 businesses now spend more than $1 million annually on Claude. That number doubled from 500 in under two months.

Claude is the only frontier model available on all three major clouds: AWS Bedrock, Google Vertex AI, and Microsoft Azure Foundry. For an enterprise procurement team, accessing Claude through their existing cloud account removes the single biggest friction point in adoption. Nobody has to sign a new vendor agreement. They just turn it on.

Anthropic shipped 50+ features in 52 days in early 2026. Claude for Excel. Claude for PowerPoint. Cowork. Healthcare integrations. Live financial data feeds from Moody’s, S&P Global, and the London Stock Exchange. Their own engineers use Claude for about 60% of their work and push 60-100 internal releases per day. The product is building the product.

Two things happened during this period that belong in the timeline, not in a separate critique section. In early 2026, Anthropic revised its flagship safety pledge, replacing a categorical prohibition on unsafe training with a more flexible framework. Time reported the change came as the company was riding “technological and commercial successes.”

And in February 2026, Anthropic disclosed that three Chinese AI labs had used 24,000 fake accounts to extract over 16 million conversations from Claude. DeepSeek used the outputs to generate censorship-safe alternatives to politically sensitive queries.

The Pentagon Stand (February-March 2026)

In July 2025, Anthropic signed a $200 million DoD contract through Palantir with two restrictions: no mass domestic surveillance, no fully autonomous weapons. In January 2026, Defense Secretary Hegseth issued a memo requiring “any lawful use” language on all AI contracts. On February 24, he gave Amodei until 5:01 PM Friday to remove the restrictions.

Amodei refused. “We cannot in good conscience accede to their request.”

The Pentagon gave the contract to OpenAI hours later, designated Anthropic a “supply chain risk”, and Trump directed all federal agencies to stop using Anthropic products. Then ChatGPT uninstalls surged 295%. Claude rose to #1 on the U.S. App Store. People wrote “Thank you” in chalk on the sidewalk outside Anthropic’s office. Altman later admitted the Pentagon deal “looked opportunistic and sloppy.”

Here’s where it gets uncomfortable. While the Pentagon was banning Anthropic, Claude was actively being used in U.S. military operations in Iran through Palantir’s Maven Smart System. The U.S. struck over 1,000 targets in the first 24 hours with AI assistance. The Pentagon banned the company whose AI it couldn’t stop using.

On March 26, Judge Rita Lin issued a 43-page ruling calling the supply chain designation “classic illegal First Amendment retaliation.” Digiday argued the $200 million Anthropic walked away from “may turn out to be the best marketing spend in Silicon Valley for years.” But Anthropic’s CFO estimated the designation could cost “multiple billions” in 2026 revenue. Calling it a brand play only works in hindsight.

The Revenue Staircase

No company experienced growth like that.

Late 2023: ~$100M. December 2024: $1B. May 2025: ~$3B. August 2025: Over $5B. December 2025: ~$9B. February 2026: $14B. March 2026: $19B. April 2026: Over $30B.

That’s $11 billion in ARR added in one month. Epoch AI calculates Anthropic has been growing at roughly 10x per year versus OpenAI’s 3.4x.

One caveat. Anthropic reports cloud marketplace revenue on a gross basis: when a customer buys Claude through AWS, the full billing shows up as Anthropic revenue, with Amazon’s cut expensed separately. OpenAI reports Azure revenue net. Per OfficeChai: Anthropic has “probably” surpassed OpenAI in gross terms, “probably not yet” in net. When the S-1 drops, this will be the most scrutinized number in the document.

That growth also comes with significant investments. Dario Amodei said this on the Dwarkesh Patel podcast when talking about scaling and investment costs: “If I’m just off by a year in that rate of growth, or if the growth rate is 5 times a year instead of 10 times a year, then you go bankrupt.”

The Money Behind the Machine

May 2021: $124M Series A. April 2022: $580M Series B (SBF). May 2023: $450M at $4.1B Series C. 2023-2024: Amazon’s $8B across three tranches. March 2025: $3.5B at $61.5B Series E. September 2025: $13B at $183B Series F. February 2026: $30B at $380B Series G.

Total raised: roughly $67 billion. Valuation: $4.1 billion to $380 billion in 34 months. That’s 93x.

The burn rate: roughly $19 billion in 2026 on training and inference infrastructure. Gross margins around 40%, well below the 77% that typically justifies software valuations. $80 billion committed in cloud costs through 2029. The Register headlined the Series G as “Investors toss another $30B atop the Anthropic money furnace.” Unkind, but not wrong. Anthropic projects break-even by 2028.

What’s Next

Most of what follows leaked in the last two weeks.

Claude Mythos. On March 26, a CMS misconfiguration exposed nearly 3,000 unpublished documents, including draft blog posts describing a new model tier above Opus. Internal codename: Capybara. Anthropic confirmed it’s real and called it “a step change and the most capable we’ve built to date.” The leaked drafts claim it is “far ahead of any other AI model in cyber capabilities.” Anthropic privately warned government officials it makes large-scale cyberattacks “significantly more likely in 2026.” Cybersecurity stocks fell 7-11%. The model that’s supposed to be the next revenue engine is, by its creator’s own assessment, also a weapon.

KAIROS. Five days later, Claude Code’s entire source code leaked via npm. 512,000 lines across 1,900 files. 84,000 GitHub stars before takedown, 8,000+ mirrors. The headline discovery: KAIROS, mentioned 150+ times, is a fully built autonomous daemon that polls itself every few seconds asking “anything worth doing right now?” It runs persistently, manages sessions, respawns on failure, and includes autoDream, a companion that consolidates the agent’s memory while you sleep. Built and waiting behind a feature flag. The leak also revealed anti-distillation controls that inject fake tool definitions to poison competitors. Two major leaks in five days is a pattern, not a fluke.

IPO. Anthropic is reportedly targeting October 2026, with Goldman Sachs, JPMorgan, and Morgan Stanley competing to underwrite. Target raise: over $60 billion.

The competition. OpenAI just closed $122 billion at an $852 billion valuation. Both companies are racing to IPO first.

The $70 billion question. Anthropic’s internal projections target $70 billion in revenue by 2028. Getting there means Mythos needs to ship and KAIROS needs to launch. The enterprise flywheel needs to keep compounding. Gross margins need to climb from 40% toward 77%. And a company that leaked two major assets in five days needs to stop leaking. All of it hinges on things that haven’t happened yet.

Key Takeaways

1. Anthropic grew from $1B to $30B ARR in 16 months by winning enterprise customers (85% of revenue), building Claude Code into a $2.5B product in six months, and being the only frontier model on all three major clouds. No enterprise software company has ever compounded this fast.

2. Claude Code is the accidental center of gravity. A product that started as one engineer giving Claude access to a filesystem now writes 4% of all GitHub commits, generates roughly a fifth of Anthropic’s revenue, and was adopted by Microsoft’s own engineering teams over their own product.

3. The Pentagon showdown was the defining brand moment in AI so far. Anthropic walked away from $200M, got banned by the federal government, won a First Amendment case, and had its AI used in a war anyway. The story doesn’t resolve neatly.

4. The $30B is real but fragile. 40% gross margins, a CEO who openly acknowledges the bankruptcy risk of slowing growth, a safety pledge loosened under commercial pressure, and two major security leaks in five days. The revenue is extraordinary. The durability is anyone’s guess.

— Alex Hrankin

Subscribe free — AI Hot takes, news, insights and deep dives.

Last Tuesday I was pulling together research for an article. Six agents running in parallel, each scraping different categories of sources, saving structured notes into folders on my hard drive. The whole collection phase took about forty minutes. A year ago that was a full day of tabbing between Google, copying quotes into documents, losing my place, starting over.

I didn’t do any of this in a chat window. I didn’t copy and paste a single thing. The AI was working directly in my files, on my computer, creating real documents I could open in Finder when it was done.

Most people still think Claude is a chatbot. You type a question, get an answer. Maybe upload a PDF. And for most of 2024, that mental model was basically right.

It’s not right anymore. The gap between “chatbot” and what’s actually possible has gotten so wide that most professionals are leaving an absurd amount of value on the table without knowing it.

This is about Claude Cowork. But to understand why it matters, I need to tell you about Claude Code first.

The Backstory Nobody Told You

Claude Code launched as a developer tool. Terminal-based, command-line, the kind of thing where you type instructions into a black screen and code happens. Developers loved it. Nobody outside of engineering paid attention.

Then something unexpected happened. Developers started using it for things that had nothing to do with code. File organization. Data analysis. Research synthesis. Building dashboards. Processing invoices. One agency used it for sales, marketing, and delivery by switching between project folders.

Why? Because Claude Code had three things the chatbot didn’t: it could read and write your actual files, it could run Python to analyze data and verify results programmatically, and it could spin up sub-agents that worked in parallel on complex tasks for extended periods. That’s a fundamentally different thing from a text box that generates a response and forgets you exist.

I started using it for coding and realized within a few weeks that it was more useful for everything else. Nine months later, I’ve used it for market research, competitive analysis, content production, data reconciliation, presentation drafts, and building this entire newsletter’s pipeline from scratch. The word “Code” in the name scared people off. The tool was never limited to what the name suggested.

Anthropic noticed. According to VentureBeat, their Head of Americas said: “In 2025 Claude transformed how developers work; in 2026 it will do the same for knowledge work.” So they built Cowork.

What Cowork Actually Is

Cowork is Claude Code wrapped in a native desktop application. Same agent engine underneath. Same capabilities. But instead of a terminal, you get a graphical interface with buttons and a sidebar and a file browser. It’s the third tab in Claude Desktop, sitting right next to Chat and Code.

Calling it “Claude Code with a nicer UI” undersells what changed.

File system access. You point Cowork at a folder on your computer and it can read, create, and modify files inside that folder. No uploading. No downloading. No copy-paste loop. It works directly in your actual file structure, the same way you would if you opened Finder and started organizing things yourself. Cowork runs in an isolated virtual machine on your computer, so your files stay local and nothing gets sent to Anthropic for training.

Connectors. This is where it gets interesting. Cowork can plug into your other tools via MCP (Model Context Protocol, the universal standard I covered in AI Curious to AI Capable). Google Drive, Gmail, Calendar, Slack, Notion, Linear, Jira, Figma, GitHub, Stripe, and dozens more. You set permissions per connector: allow everything, ask before acting, or block entirely. Connect your meeting notes tool and Cowork reads what was discussed in your last product review, then edits the prototype sitting in your project folder based on the feedback. One workflow, no tab switching.

Instructions. You write global instructions that apply to every session (”I work in fintech, always format currency with two decimal places, my reports go to a CFO who skims”) and folder-level instructions that activate when Cowork opens a specific project. This is how Cowork remembers who you are and how you work, without you repeating yourself every time. Think of it as the onboarding document that follows every new hire, except this one actually gets read.

Plugins. Pre-built bundles of skills, connectors, and instructions packaged for specific roles. Anthropic ships 11 official plugins covering Productivity, Sales, Marketing, Legal, Finance, Product Management, Data Analysis, and more. The Sales plugin connects your CRM and knows how to draft proposals. The Legal plugin can analyze contracts. And you can build your own, which I’ll get to.

Two features that deserve special attention:

Plan Mode. Before Cowork touches anything, it shows you a plan. What it’s going to do, which files it will read, what it will create. You approve before it executes. I think this is the single most underrated feature in any AI tool right now. On complex tasks, having Cowork plan first instead of trial-and-erroring through execution cuts task time by roughly two-thirds and produces better results. The first few times I skipped planning to save time, I ended up redoing the entire task from scratch. I stopped skipping.

AskUserQuestion. Mid-task, Cowork can pause and ask you clarifying questions with multiple-choice options before continuing. Instead of you trying to write the perfect instruction up front and hoping it guesses correctly, the AI tells you what it’s uncertain about and asks you to decide. This inverts the normal AI interaction, and the difference in output quality is immediate.

And then there’s Scheduled Tasks, which just launched around February 25. You can set Cowork to run tasks automatically on a recurring cadence: hourly, daily, weekly, or weekdays only. A morning briefing that summarizes your Slack, email, and calendar from the past 24 hours. A weekly report that pulls data from Google Drive into a formatted summary. Recurring competitor monitoring. The first time a scheduled task runs, Claude actually rewrites the underlying instructions based on what it learned, so subsequent runs get faster and more targeted.

One real limitation: scheduled tasks only run while your computer is awake and Claude Desktop is open. If the app is closed when a task is due, it skips and catches up when you reopen. Not ideal, but the feature is days old.

What It’s Actually Useful For

I could list abstract capabilities all day. Here’s what people are actually doing.

Research that produces real deliverables. Not “here’s a summary in the chat window.” I mean: you tell Cowork to research a topic, and forty minutes later there’s an Excel comparison matrix, a written report, and a rough presentation draft sitting in your project folder. Multiple files, cross-referenced, with actual data. I do this weekly for this newsletter.

Data analysis across messy real-world files. Bank statements from three different banks, each with different column names and date formats. Cowork reads all of them, writes Python to normalize and merge the data, verifies the totals programmatically, and flags discrepancies. One practitioner on Hacker News described a 3-day resume processing task (cross-referencing accreditation databases, salary benchmarks, and publication indexes across roughly 100 resumes) done in 30 minutes.

Invoice and expense processing. Drop a folder of invoices into your project directory. Cowork extracts vendor names, amounts, due dates, categorizes everything, outputs a clean spreadsheet. One user reported processing hundreds of Excel files for tax filing by extracting rows without VAT across all of them and aggregating into one document. Another pointed Cowork at random expense screenshots in their Downloads folder and had a working budget tracker with formulas in about four minutes.

The use cases that surprised me weren’t the data-heavy ones. It was the mundane stuff. Meeting prep, for instance: connect your calendar and meeting notes tool, and before a client call Cowork pulls the company’s recent news, LinkedIn profiles of attendees, their pain points based on previous conversations, and compiles it into a briefing doc. The kind of prep you know you should do but never have time for. I used to guilt myself about showing up to calls underprepared. Now I don’t.

File organization at scale. One reviewer tested Cowork on 500 Google Drive files and it created logical folder structures, renamed with consistent conventions, and flagged duplicates in under 10 minutes. I use it for sorting messy invoice folders, but the use case extends to any situation where you have hundreds of files that need structure.

Local reporting dashboards without any engineering involvement. Product usage data, Google Ads performance, Meta stats. Cowork reads the raw data files, writes Python to build visualizations, and saves them locally. One team used it to pull a specific click count from a PostHog dashboard without filing a single ticket.

And there’s a long tail of creative uses: analyzing 320 podcast transcripts to extract the 10 most important themes in 15 minutes. Building custom micro-apps. Sprint retrospective analysis across six months of notes in different formats. Creating SOPs from scattered documents and meeting transcripts.

The mental shift is thinking of Cowork as asynchronous delegation, not interactive chat. You hand it a task, walk away, come back to finished work. Every.to put it well: once you experience handing something off and finding it done an hour later, something clicks.

How to Set This Up Properly

This is where nine months of using Claude Code pays off. The difference between Cowork being a gimmick and Cowork being genuinely useful comes down to how you set up your environment. I’ve gotten this wrong enough times to know what works.

Create a dedicated project folder. Make a folder called something like “Cowork-[YourCompany]” or “Cowork-[Project]”. This is the folder you’ll point Cowork at. Never point it at your entire Documents folder or your home directory. Start contained, expand later.

Inside that folder, add a file called CLAUDE.md. This is the project memory file. When Cowork opens this folder, it reads CLAUDE.md first and carries that context through the entire session. Write a few paragraphs: what your company does, what your role is, what tools you use, what formats you prefer, any conventions or terminology specific to your work.

Mine has about 300 lines. It covers our brand, our content formats, our Notion database structure, our publishing schedule, even specific writing rules for different article types. When I start a new session, Cowork already knows all of this. No re-explaining. (Cowork also has a visual Instructions panel that does the same thing, so you can use either approach. I prefer the markdown file because I can version-control it.)

Organize knowledge in subfolders. Create directories for invoices, research, presentations, company knowledge, client materials. The better organized your folder structure, the faster Cowork cross-references information across documents. If everything is dumped in one flat folder, Cowork has to read every file to find what it needs. Structure is context.

Use markdown for everything you can. This might seem like a preference, but it’s actually practical. Markdown files use dramatically fewer tokens than Word documents or PDFs, which means Cowork can hold more of your content in its working memory simultaneously. Start your reports, notes, and drafts in markdown. Convert to PDF or Word at the end when you need a polished deliverable. I write this entire newsletter in markdown. Every research document, every article draft, every project plan.

I spent an embarrassing amount of time early on fighting with file formats before figuring this out. Uploading a 40-page Word doc when a 40-page markdown file would’ve been a fraction of the token cost. Match the format to the consumer. When the consumer is AI, markdown wins.

Connect everything you can. Go to the connectors panel and link your tools: Google Drive, Slack, your calendar, your CRM, whatever you use daily. The value compounds with each connection because Cowork can pull context from one tool while working in another.

My favorite example: I connected Granola (my meeting notes app). Now Cowork reads the transcript from my latest product review meeting, identifies the feedback on a specific feature, opens the prototype file in my project folder, and makes the edits. That used to be three separate workflows. Set granular permissions for each connector. Allow read access broadly, restrict write access to things you’re comfortable with, block actions you never want automated. Worth spending ten minutes getting right.

Use Plan Mode for anything complex. Tell Cowork to plan first: “Research X and create a plan before executing. Show me the plan and wait for my approval.” Cowork researches, thinks through the approach, and presents a structured plan. You review, adjust, approve.

I do this for everything. It sounds slower. It’s dramatically faster. Practitioners consistently report that planned execution takes roughly a third of the time that trial-and-error execution takes, and the output quality is noticeably higher because Cowork has already thought through the edge cases before touching anything.

Start with the official plugins, then build your own. The 11 Anthropic plugins are free and open-source. Install the one closest to your role. Use them for a week to see how skills and commands work in practice.

Then try building a custom one. The barrier is lower than you’d expect. Plugins are just files: a JSON manifest that describes what the plugin does, plus markdown files containing the instructions for each skill. You don’t need to write any code. You can ask Cowork to build a plugin for you by describing what you want it to do. One practitioner replaced a multi-step article editing workflow with a single /polish-article command that runs polishing, structuring, and SEO optimization in one pass.

The distinction: a skill is a single reusable instruction set (like a recipe). A plugin bundles multiple skills, connectors, and configurations into one installable package (like a cookbook). Start with individual skills. Bundle them into plugins once you know what you need.

Set up scheduled tasks for recurring work. Open Cowork, type /schedule, describe the recurring task, set the cadence. Some ideas that work well:

• Daily morning briefing: summarize Slack, email, and calendar from the past 24 hours

• Weekly report compilation: pull data from spreadsheets or Drive into a formatted summary

• Recurring file organization: sort incoming invoices or documents on a weekly schedule

• Competitor monitoring: track specific companies or topics and surface changes

Your computer needs to be awake and Claude Desktop open for scheduled tasks to run. If you close the app, it catches up when you reopen.

The Honest Caveats

I’d be doing you a disservice if I didn’t cover the rough edges. And honestly, some of these frustrated me enough that I want you to know about them before you hit them yourself.

Usage limits are real. Cowork burns through your quota faster than regular chat. A single complex task can equal dozens of normal conversations. The $20/month Pro plan gives you access, but heavy users will hit limits. The $100/month Max tier gives roughly 5x the capacity. I’m on Max and I still occasionally ration usage on intensive research days.

It doesn’t remember between sessions. Cowork starts fresh every time. That’s why the CLAUDE.md file and folder structure matter so much. Your persistent context lives in your files, not in Cowork’s memory. Some practitioners add a line to their global instructions telling Cowork to append anything valuable it discovers to a memory.md file, creating cross-session persistence at near-zero cost.

Vague instructions produce vague results. “Clean up my data” won’t work. “Scan the Sales_2025 folder, find rows with missing email addresses, highlight them in red, and create a summary of how many are missing per month” will. The specificity of your instructions is the biggest factor in output quality. This is the part that trips up most new users, and it’s the part that gets dramatically easier with practice.

And there are alternatives worth knowing about. If your company runs entirely on Google Workspace, Gemini’s in-app integration may be more useful since it’s embedded directly in Gmail, Docs, and Sheets. If you need web browsing and form-filling automation, ChatGPT’s Agent mode or Operator handle that better. And if your company lives in Microsoft 365, Copilot has native context that no external tool can match.

Cowork’s strength is local file manipulation, multi-step data processing, and long-running autonomous tasks in your own environment. If that describes your bottleneck, nothing else comes close right now.

Try It Yourself

1. Download Claude Desktop and sign up for at least Claude Pro ($20/month)

2. Create a project folder with subfolders for your main work categories

3. Add a CLAUDE.md file describing your company, role, and working preferences

4. Open the Cowork tab, point it at your folder, and start with something concrete: “Organize the files in [subfolder] into logical categories and create a summary of what you did”

5. Before your next meeting, try: “Research [company name] and [attendee names], then create a one-page briefing document in this folder”

6. Install one of the official plugins closest to your role

7. Set up your first scheduled task: /schedule + “Every weekday morning, summarize my email and calendar for the day”

8. Connect at least two tools via connectors (Google Drive and Slack are a good start)

Start small. The compounding happens fast.

— Alex

P.S. I’ve been saying for months that the word “Code” in Claude Code was the biggest branding miss in AI. The number of non-developers who never tried it because of the name is staggering. Cowork fixes that. But it also means most people are about to discover capabilities that power users have had access to for nearly a year. If you’re reading this before that wave hits, you have a head start. Use it.

The AI Bridge — AI education for professionals who want to actually use AI, not just read about it.

Subscribe free — The Signal every Tuesday, The Deep End every Saturday, Field Notes throughout the week.

I believe AI will replace many entry level jobs and automate numerous tasks (and is already doing that), and Skills is how you teach your AI employees, the same way as you would teach a human. This article is about what skills are and how they work.

Most of the work with AI in the last few years (and quite dominant still today) is a regular chat process - asking AI to do something, explain it, repeat. An important question people were asking is - how do I teach AI to do the same things I can, or the same processes the my company does? First asnwer was heavy prompting and for some - fine-tunning (which is quite expensive and only large organisations would actually do it at scale). After that emerged a new layer - workflow automations. Companies like Zapier or N8N, legacy workflow automation platforms, realised they have the foundation to embed AI into workflows and enable users to have repeated workflows. The latest answer is Skills, which emerged at the end of 2025 and introducted by Anthropic (as many innovations in AI).

What a Skill Actually Is

A skill is a set of instructions / capabilities you write once that your AI agent applies automatically whenever it’s relevant. You don’t paste it into a chat window. You don’t have to remember to invoke it. The agent reads your request, recognizes that it matches a skill you’ve created, and pulls in those instructions on its own.

In Claude, a skill lives in a folder containing a file called SKILL.md. That file has a name, a description, and whatever instructions you want the agent to follow. The description is the important part. When you ask Claude to do something, it compares your request against the descriptions of every available skill and activates the ones that match. The way it works is Claude pre-loads 1-2 sentence summaries of each skill in it’s own context, thus it “knows” which skills are available. This also allows to optimise the usage, as your client report skill template doesn’t sit in memory fully. It loads when you actually ask for a report.

But here’s where people get it wrong. A common misconception is that skills are “just markdown files.” A skill is a folder. That folder can include scripts, reference documents, templates, example outputs, data files. The SKILL.md is the entry point, but the whole directory is the skill. Anthropic’s documentation makes this clear, and the distinction matters more than you’d think.

A skill that includes a validation script it can run, or a template file it can copy and fill in, is a fundamentally different thing from a skill that’s just a page of text.

Example: one of my clients (Hedge Fund) had a workflow of analysing the latest barclays daily report, Bloomberg terminal data and creating an internal one-pager report which is sent to clients. This is a hour or more of work every day of an analyst. We created a skill with detailed instructions to get Bloomberg data via API, parse Barclays report into csv, python scripts to analyse the data, and than create a one-pager from a template with detailed examples of past reports, as well as do’s and don’ts. Now they just have to run the skills and in 10-20 minutes they have a proper one pager with validation (we make AI check all of the data behind the scenes to avoid halucinations).

Skills aren’t limited to Claude. They follow the Agent Skills open standard, which works across multiple AI tools. A skill you build for one agent should, in theory, work in another. We’re early on that cross-tool compatibility, but the direction is clear.

Eight Types of Skills (And What They Look Like in Practice)

Anthropic’s engineering team has been using skills internally for months, with hundreds in active use across the company. After looking at how they categorize their skills and thinking about how the concept applies beyond software engineering, I’ve landed on eight types that cover what most professionals actually need. Whether you’re running a hedge fund, managing a plumbing business, or consulting for Fortune 500 clients, the same categories apply.

1. Domain Knowledge & Standards. Every business has internal knowledge that AI doesn’t have by default. An accounting firm’s chart of accounts with its edge cases (meals with clients = 6410, meals for team events = 6420, meals during travel = 6450). A consulting firm’s specific frameworks. A hedge fund’s sector playbooks with the metrics that matter for each industry. A plumbing company’s service area, pricing tiers, and warranty policies. You teach the AI your world once, and it stops making assumptions that don’t fit.

2. Research & Investigation. This covers both proactive research (a consultant sizing a market, a financial analyst building a comp table) and reactive diagnosis (an accountant figuring out why two reports don’t tie, an analyst investigating why a stock moved overnight). A strategy consultant might have a skill that takes a client’s question like “Should we enter the EU market?” and builds a hypothesis tree with testable sub-hypotheses. A hedge fund analyst might have one that pulls every possible explanation for a price move: earnings, analyst changes, insider activity, sector rotation, options flow.

3. Data & Analysis. Connect to your data and make sense of it. An accounting manager’s variance analysis skill that pulls actuals versus budget, flags anything over 10%, and drafts a one-paragraph explanation for each variance. A car dealership’s per-unit profit calculator that tracks purchase price plus reconditioning plus floor plan interest versus sale price. A portfolio exposure dashboard that shows daily P&L by position, sector exposure, and top contributors. Instead of explaining your data architecture every time you need a metric, you explain it once.

4. Document Templates & Generation. Every business creates standardized documents. An investment memo in the fund’s specific format (thesis in 3 sentences, key catalysts with timing, bear case, valuation with 3 scenarios). A consulting firm’s slide deck where every headline states the insight, not the topic (”The market grew 12% but all growth came from one segment” instead of “Market Overview”). Engagement letters. Listing descriptions for a real estate agent. Job postings for an HR manager. The skill knows your format, your tone, your conventions.

5. Communication & Drafting. This might be the highest-volume category for most professionals, and it’s the one people overlook. A plumber’s Google review responder that drafts personalized replies (5-star reviews get a thank-you mentioning the specific job, 1-star reviews get a professional recovery response). A car dealership’s follow-up drafter that references the exact car the customer test-drove and the concern they raised. An HR manager’s offer letter writer that knows the company’s tone, legal requirements, and salary band conventions. A hedge fund analyst’s earnings quick-take that pulls the numbers within minutes and drafts a 5-line summary the PM can read in 30 seconds.

6. Review & Compliance. Check work against quality and regulatory standards. An accounting manager’s month-end close checklist that verifies every journal entry posted, confirms bank reconciliations, and flags anything still open. A lawyer’s contract review skill that flags unusual indemnification clauses, non-standard termination provisions, and missing limitation of liability. A consultant’s QA skill that catches the things that make a partner send the deck back at 11 PM: unsourced numbers, mismatched charts, leftover “TBD” placeholders.

7. Process Automation. Execute repetitive multi-step workflows with one command. A weekly recap that compiles activity across your tools into a formatted summary. An invoice processor that takes vendor invoices (PDF, email, photo), extracts the key data, matches to the right account, and outputs a batch entry file. A morning research digest that pulls overnight news and price moves for every stock in the portfolio. I have a skill that generates five Substack Notes from a published article. The interesting thing about these skills is they get better over time: save logs of previous runs and the agent starts comparing against its own history.

8. Monitoring & Alerts. Passive, continuous watching for changes that need your attention. A hedge fund’s SEC filing monitor that flags new 13F filings and insider transactions. A car dealership’s competitor price tracker that alerts when a nearby dealer lists the same model significantly cheaper. An accounting manager’s AP aging skill that identifies chronically late payers and drafts escalation emails. A plumber’s parts reorder alert that flags when common inventory drops below threshold. These are distinct from process automation because they’re always-on and event-driven rather than manually triggered.

For engineering and development teams: Anthropic’s internal taxonomy includes two additional categories that are technical in nature — CI/CD & Deployment (skills that build, test, and deploy code, like a “babysit PR” skill that monitors pull requests and auto-merges) and Infrastructure Operations (routine maintenance with guardrails, like finding orphaned cloud resources). These use the same skill structure but apply specifically to software workflows.

I keep going back to this list when I’m between tasks, trying to figure out what to automate next. Most people I talk to can immediately identify three or four categories that apply to their work.

Where Skills Live (And Why It Matters)

Where you store a skill determines who can use it, and the answer changes the entire dynamic of how teams work with AI.

Personal skills go in your home directory (~/.claude/skills/). These follow you across every project. Your writing voice. Your preferred report format. How you like data presented. These are your preferences, portable everywhere you go.

Project skills go in .claude/skills/ inside your project directory. Anyone who works from that project gets them automatically. No extra setup. This is where team standards live — your firm’s review checklist, your company’s brand guidelines, your department’s reporting conventions. And because they’re files in a shared directory, updating a skill is as simple as editing the file. Everyone on the team gets the update immediately.

Enterprise skills get deployed organization-wide through managed settings and take the highest priority. If your company has an enterprise “client-report” skill and you create a personal one with the same name, the enterprise version wins. Always. This is intentional. When the word is “must” (must follow these compliance procedures, must use this disclosure language, must format reports this way), enterprise skills are the mechanism.

Plugin skills come from installable packages distributed through marketplaces. Best for functionality that isn’t project-specific and could benefit a broader community.

The priority hierarchy is strict: Enterprise beats Personal beats Project beats Plugins. So if you’re ever confused about why a skill isn’t firing, check whether something higher in the chain is shadowing it.

Skills vs. Everything Else

Claude has several ways to customize behavior, and I’ve seen people shove the wrong thing into the wrong feature and wonder why it feels clunky. Getting this right once saves you from building the wrong thing repeatedly.

Project Memory loads into every single conversation. It’s always-on project memory. If you want Claude to always follow your brand voice, never touch certain sensitive files, or remember that your fiscal year starts in April, that goes in CLAUDE.md. The downside: everything in that file consumes context in every session, whether it’s relevant or not.

Skills load on demand. Only the name and description sit in memory. The full instructions only load when Claude recognizes a matching request. Your fifty-point client report checklist doesn’t eat context when you’re drafting an email. It appears when you actually ask for a report. This on-demand loading is the key difference, and it’s why skills scale better than cramming everything into CLAUDE.md.

Hooks fire on events. A hook might run a quality check every time Claude saves a file, or validate input before certain actions. They’re automatic, but they’re triggered by actions, not by what you’re asking about. Skills respond to intent. Hooks respond to events.

Subagents run in isolated contexts. They receive a task, work on it independently, and return results. Here’s something that catches people off guard: subagents don’t automatically see your skills. If you delegate a task to a built-in subagent, it can’t access your skills at all. Only custom subagents you explicitly configure can use them. This is a design choice, not a bug. But it surprises nearly everyone the first time.

MCP servers provide external tool connections. A different category entirely. Think of MCP as the plumbing that connects your AI to Slack, Google Drive, your CRM, your accounting software. Skills tell the agent what to do. MCP gives the agent access to do it.

Use all of them together. A typical setup: Project memory / instructions for always-on standards, skills for task-specific knowledge, hooks for automated quality checks, subagents for delegated parallel work, MCP for external connections. Each handles its specialty.

Building Your First Skill

Claude (or other AI system) can help you create a skill v1 for you. Which should be more than enough to get you started, although creating better and more custom skills takes effort, almost like training an employee. This is why the section below on how skills are set up technically is useful to know.

Technical deep-dive: Create a directory inside your skills folder. Give it a descriptive name. Inside, create a SKILL.md file with two parts: YAML frontmatter (between --- markers) that tells Claude about the skill, and markdown content below with the actual instructions.

The frontmatter has two fields that matter most: name (what the skill is called) and description (when Claude should use it). The description is everything. Claude uses it to decide whether your skill matches an incoming request, and Anthropic’s course on skills makes a point I think about often: if someone told you “your job is to help with docs,” you wouldn’t know what to do. Claude thinks the same way. A good description answers two questions: what does this skill do, and when should Claude use it?

Beyond those two, there are optional fields that add real power. allowed-tools restricts which tools Claude can use when the skill is active — useful when you want a skill that can read your files but never modify them. disable-model-invocation: true prevents Claude from triggering the skill automatically, so only you can invoke it manually. You want this for anything with real consequences: sending client emails, posting to social media, submitting financial data. And model lets you specify which Claude model runs the skill.

And then there are supporting files. Templates that Claude copies and fills in. Reference docs it reads only when needed. Scripts it can execute. The Anthropic engineering team recommends thinking of the entire file system as a form of progressive disclosure. Tell Claude what files exist in your skill directory, and it reads them at the right time. Keep SKILL.md under 500 lines. Move detailed reference material into separate files. Claude loads them only when the task requires it, which keeps your context window clean.

One pattern I keep coming back to: skills with memory. A standup skill that appends every post it writes to a log file. Next time it runs, Claude reads its own history and knows what changed since yesterday. I didn’t expect this to work as well as it does. The skill genuinely gets smarter each time it runs, without you doing anything.

The Tips That Actually Matter

I recommend to everyone to read Anthropic’s internal lessons on building skills. But here’s the three things stand out as genuinely non-obvious:

Don’t state the obvious. Claude already knows a lot about general business practices, common formats, and standard tools. A skill that repeats what Claude already knows is wasting context. Focus on the information that pushes the agent out of its default behavior. What are the specific quirks of your business? Where does Claude consistently make the wrong assumption about your industry? That’s what belongs in a skill. An accounting firm’s skill shouldn’t explain what a balance sheet is. It should explain that your firm uses a non-standard chart of accounts for real estate clients and here’s why.

Build a gotchas section. The Anthropic team says the highest-signal content in any skill is the gotchas section. Things Claude gets wrong. Edge cases it misses. Assumptions it makes that don’t hold in your business. Start with a few. Add more every time Claude hits a new failure. A car dealership’s listing skill might start with one gotcha: “never say ‘like new’ in a listing — it’s a legal liability.” A month later it has twelve gotchas, and every listing comes out clean. There’s something satisfying about this. Every mistake becomes an investment in a system that won’t make that mistake again.

The description field is for the model, not for you. When Claude Code starts a session, it builds a listing of every available skill with its description. This is what Claude scans to decide “is there a skill for this request?” The description isn’t a summary for humans. It’s a trigger specification for the AI. Include the phrases people actually use when they’d want this skill to activate. If your skill isn’t triggering when expected, the fix is almost always a better description.

And one thing I’ve learned from my own use: avoid railroading. Because skills are reusable across many different situations, being too specific can backfire. A skill that says “always output exactly seven bullet points” will produce garbage when the task needs three or fifteen. Give Claude the information it needs, but leave room for it to adapt.

Sharing and Scaling

A skill that only you use is helpful. The same skill shared across your team standardizes how everyone works.

The simplest path: save your skills in a shared project directory under .claude/skills/ if you’re using Claude Code or add them as a capability in Claude Desktop app. Anyone on the team gets them. This works well for small teams.

For broader distribution, there are plugin marketplaces. Package your skills into installable bundles and other users can discover and install them. Anthropic’s internal approach is worth knowing: they don’t have a centralized team deciding which skills go in their marketplace. Useful skills surface organically. Someone builds a skill, points people to it in Slack, and once it gets traction, it moves into the marketplace. But they also warn that it’s easy to create bad or redundant skills, so curation before release matters.

For enterprises, managed settings deploy skills organization-wide with the highest priority. Mandatory compliance workflows. Regulatory language requirements. Reporting standards that must be consistent across every team. The “must” is doing real work there.

One more thing about scaling: measuring. Anthropic uses hooks to log which skills get used and how often, so they can find skills that are popular and skills that are under-triggering relative to expectations. If you’re deploying skills across an organization, knowing what’s actually being used (versus what you thought would be used) is the difference between a growing system and a graveyard of good intentions.

Key Takeaways

1. A skill is a folder of instructions that your AI agent applies automatically when relevant. You write them once. The agent remembers them forever. No copying, no pasting, no repeating yourself.

2. Skills load on demand, not all at once. Only the name and description sit in memory until the agent recognizes a matching request. This is why skills scale better than stuffing everything into a single project memory file.

3. The eight skill categories are a practical checklist, not an academic taxonomy. Scan the list, find the categories that match your repetitive work, and start there. Communication & Drafting and Process Automation are the highest-volume categories for most professionals.

What to Do This Week

Start with the task you explain to Claude most often. The one where you catch yourself mid-instruction thinking “I’ve said this exact thing before.” That’s your first skill.

Write the skill. Keep the instructions short. Add a gotchas section, even if it only has one item. Test it by asking Claude something that should trigger it. If it doesn’t trigger, improve the description. Add the phrases you actually use.

Then look at the eight categories. Which ones apply to your work? Most people I talk to can immediately identify three or four skills they should have built weeks ago. The friction isn’t technical. It’s just knowing that the option exists.

I spent three months using Claude before I understood skills properly. During that time I typed the same instructions hundreds of times, each time thinking I was being efficient because the output was good. The output was fine. My process was embarrassing. Every skill you build is a set of instructions you’ll never type again, and that time adds up faster than you’d think.

— Alex

TL;DR: If you’re still typing the same instructions into your AI agent every week, you’re working harder than you need to. Skills fix that.

The AI Bridge — AI education for professionals who want to actually use AI, not just read about it.

Subscribe free — The Signal every Tuesday, The Deep End every Saturday, Field Notes throughout the week.

“Emails, mostly. Sometimes I’ll ask it to clean up a deck.”

That’s it. Twenty dollars a month to autocorrect. And she’s not behind because she’s incurious or lazy. She’s behind because nobody told her what happened. Nobody explained that AI stopped being a chatbot about fourteen months ago and she’s still interacting with it like it’s 2024.

I keep running into this same conversation. Smart, senior people who use ChatGPT for the occasional email draft or Perplexity when they want a sourced answer. Maybe Gemini, because Google bundled it in. The gap between where they are and where the tools actually are feels impossibly wide. Two things keep it that way: they don’t know what exists beyond the chatbot, and they assume the advanced stuff is engineer-only territory. Both of those were true eighteen months ago. Neither is true now.

The stakes aren’t abstract. PwC found that workers with AI skills earn a 56% premium over peers in the same roles without them. And IBM’s data shows 78% of enterprises have deployed AI tools while only 6% of employees feel comfortable using them. Read that again: the tools are already there. The understanding isn’t.

Your AI Stopped Forgetting

You know that thing where you’re talking to ChatGPT and halfway through the conversation it just... forgets what you told it? You explain your whole situation, give it all this context, and ten messages later it’s acting like you never said any of it. That happened because these models have a working memory called a context window, and two years ago that window was tiny. About six pages of text.

Now Claude holds a full novel. Gemini can hold several. You can drop a fifty-page quarterly report into a conversation and say “find the three things my CFO will care about.” That’s not a incremental improvement. That’s a category change in what AI is even for.

I wish I could tell you bigger windows just automatically make everything better. They don't. There's this thing called context rot where you throw too much in and the model gets sloppy, starts mixing up details. But the hard ceiling that used to define how you worked with AI basically disappeared.

AI Isn’t Just a Text Box Anymore

Today’s models see, hear, and read documents natively. Photograph a contract on your phone, get a plain-English risk summary in thirty seconds. Talk to Claude with your voice while commuting and have a real back-and-forth about a problem you’re working through. Upload a 200-page PDF and ask for the trends nobody mentioned in the last board meeting. For a lot of professionals, especially people who think better out loud, this is the entry point that clicks before anything else does.

AI Plugged Into Your Actual Tools

This is the change I think most people have missed entirely.

Up until pretty recently, every interaction with AI went like this: you copy something from your CRM, paste it into ChatGPT, get a response, copy that back out. Hundreds of times a week. Just shuttling text between windows. Then Anthropic built something called MCP — Model Context Protocol. Easiest way to think about it: before USB-C, every phone had a different charger cable. MCP did the same thing but for AI. One universal standard that lets any AI plug directly into any data source or tool. Over sixteen thousand integrations built in about a year and a half. OpenAI adopted it. Then Google. Then Microsoft.

What that looks like in practice: your AI reads your Slack messages directly, pulls from Google Drive, checks your calendar, updates your project management tools. No copy-paste. For this newsletter, my content pipeline runs through Notion. I have MCP set up so Claude reads the database directly, knows what’s in progress, what needs attention. I don’t open Notion for that anymore.

And the tooling is already evolving past raw connections into something called skills: pre-loaded instruction sets that tell AI not just that a tool exists, but how to use it for your specific situation. MCP gives AI access to create tasks in JIRA. A skill tells it how your team’s JIRA actually works — your naming conventions, which fields matter, what to include, what not to do. The tool access was the breakthrough. But the context about how you use those tools, the institutional knowledge that usually lives in someone’s head, is where it starts feeling less like a tool and more like a colleague who’s been onboarded.

From Chatbots to Agents

February 2025. Andrej Karpathy, co-founder of OpenAI, posts a tweet that gets four and a half million views. He’s describing how he’s building software just by talking to AI. Not writing code. Talking. He called it vibe coding. Merriam-Webster started tracking the term within weeks. (I find it slightly hilarious that the way we determine whether a technology shift is legitimate in 2026 is whether dictionary editors start paying attention. There’s probably a metaphor about language and AI in there somewhere, but I’ll leave it alone.)

Vibe coding was one step in a much bigger shift. The trajectory looks like this:

2021: AI finishes the line of code you’re typing. Smart autocomplete. 2023: You describe what you want, get code back, paste it in. Early 2025: Tools like Cursor and Bolt let you build full applications by describing them. Late 2025: AI starts working directly inside your project, reading files, running tests, checking its own work. 2026: Teams of AI agents, each with a role, coordinating on their own.

It’s like going from doing everything yourself to managing a team. You’re still deciding, still reviewing. But you’re not doing every task with your own hands. And 63% of active vibe coding users aren’t developers. They’re product managers, marketers, ops people.

Fair warning: you will be slower before you’re faster with this, the same way learning to ride a bike makes you slower than walking for the first few hours. The wobbling is the learning.

AI That Does Things, Not Just Says Things

Two flavors of this emerged almost simultaneously.

First: computer use. Some AI systems can now literally see your screen, move the mouse, click buttons, and type. Perplexity just launched Computer, a multi-model agent that coordinates 19 different AI models to run long workflows on a virtual machine in the background. You describe a goal, it decomposes it into subtasks, and works for hours while you do something else. The AI gets its own computer to work on.

Second: browser agents. AI that browses the web for you, fills out forms, moves through sites the way a human would. Claude in Chrome lets you show Claude what you’re looking at and have it take actions on your behalf. ChatGPT folded its Operator agent into a unified Agent Mode. Perplexity built an entire AI-native browser called Comet.

I find this genuinely unsettling sometimes. Last week I set an agent to research competitors for a client project and went to make coffee. Came back forty minutes later. It had made a judgment call I didn’t expect — decided one of the “competitors” I’d listed was actually a potential partner and reorganized the entire analysis around that framing. It was right. But the moment where you realize the thing made a decision without you takes getting used to. I’m still getting used to it.

Honest caveat: there’s an unsolved security problem called prompt injection, where malicious content on a web page can hijack an AI agent’s instructions. The UK’s National Cyber Security Centre warned it may never be fully eliminated. So use these agents for repetitive, low-stakes tasks. Stay in the loop on anything that matters.

The Dam Broke in February

That was the structural shift. Here’s what happened in the last two weeks alone.

Perplexity launched Computer on February 25 — the multi-model agent I mentioned above. The same week, Notion shipped Custom Agents: autonomous AI teammates that live inside your workspace and handle recurring workflows around the clock without being asked. Remote’s IT team deployed one and saved 20 hours per week, resolving a quarter of support tickets with zero human involvement. That’s a specific team at a specific company, not a demo. Then Manus went live on Telegram — a full AI agent inside a messaging app that Meta acquired for over $2 billion. And Anthropic released Claude Cowork updates, which they describe as “Claude Code for the rest of your work”: a desktop app where you give Claude access to your local folders and it handles document generation, data processing, and file management autonomously.

Four major agent products in ten days. The adoption question settled itself quietly while most people were still debating whether to try ChatGPT.

The Skill That Actually Matters

I need to be honest about something I got wrong.

When I started using AI seriously, I thought the skill was prompt engineering. Writing the perfect instruction. Getting clever with your phrasing. I had a folder of saved prompts that I organized by category. I’m embarrassed to tell you how much time I spent on that folder.

The skill isn’t prompting. It’s specificity. And there’s a difference.

People talk to AI like the genie in a lamp. You rub it, genie appears, you say “make me taller.” Genie makes you twenty feet tall. Technically correct. Terrible outcome. AI works exactly like that. It does what you ask, not what you mean.

“Professional and polished” tells AI almost nothing. “Three bullet points, each under 20 words, written for a VP who skims” tells it everything. Format, length, audience, tone, all in one sentence. The adjectives you’d normally use to describe what you want — “good,” “professional,” “high-quality” — are exactly the words AI can’t do anything useful with.

Ask it to ask you questions first. “Before you start, ask me clarifying questions until you’re confident you understand what I need.” This forces AI to surface the assumptions it would otherwise make silently. The difference between the genie guessing and the genie asking what you actually mean by “taller.”

Give it a role, then make it argue with you. “You’re a senior product strategist who’s been in B2B SaaS for fifteen years” produces fundamentally different output than a bare request. And once you get an answer you like, follow up with “now argue against this.” AI has a deep sycophancy problem — it wants to agree with you, validate your framing, tell you your idea is great. You have to actively force it into the opposing position.

Mollick points out something that I keep coming back to: the specification formats professionals already use are exactly what AI needs. PRDs. Project briefs. Consulting deliverable specs. They all share the same structure — why we’re doing this, what done looks like, what the constraints are. If you’ve ever written a decent brief for a colleague, you already know how to direct AI. You just haven’t thought of it that way yet.

I think this is going to become as fundamental a professional skill as writing a good email. Maybe more fundamental, actually, because a good email has a limited blast radius. A good AI workflow compounds every time you run it.

And it’s not just about writing better instructions. Working with AI, developing an intuition for what it can do, what it can’t, where it’ll quietly cut corners — that’s something you build over time the way you build any working relationship. You learn its tells.

Your Move

Stop treating AI like a search engine. If you’re on the free tier, upgrade. Claude Pro or ChatGPT Plus. Twenty bucks. Bring a real work problem — the learning only happens on real tasks.

Set up MCP for the tools you use most: calendar, email, CRM. Try Claude Cowork on a folder of actual work files. Try Cursor or Claude Code to build a small internal tool or prototype — you’ll be surprised what’s possible now without writing code.

Then audit your recurring work. The stuff you do every week that’s simple but manual. Ask AI if it can automate it. Ask it how. The answer is increasingly yes.

The model names will keep changing. Opus, Gemini, whatever comes next quarter. But the skill of working alongside AI, of knowing how to direct it and when to trust it, compounds in a way that doesn’t reset every time a new model drops.

— Alex

What it is, why it matters, and what 187,000 developers saw that you haven’t — yet

If you’ve been anywhere near the tech world in the last month, you’ve seen the lobster. It’s on Twitter, it’s on Discord, it’s on the front page of CNBC and Nature. It crashed GitHub’s trending charts faster than any project in the platform’s history. And behind it is a story that reads like fiction: a burned-out founder, a birthday party in Morocco, a name change war room, cryptocurrency snipers, and an AI agent that figured out how to transcribe audio messages without anyone teaching it how.

This is OpenClaw. And if you’re trying to figure out what everyone’s talking about — or whether it matters to you — here’s what I’ve found.

The Numbers That Don’t Make Sense

Before we get into the story, let’s establish the scale of what happened.

OpenClaw hit 100,000 GitHub stars in roughly two days. For context, React — the framework that powers most of the modern web — took about eight years to reach that number. Linux took twelve. OpenClaw did it over a weekend.

As of this writing: 187,000 stars. 31,500 forks. 9,384 commits across 2.5 months. Native apps for macOS, iOS, and Android. 41 messaging platform integrations. 73 built-in skills. Coverage from CNBC, Nature, TechCrunch, IBM, CrowdStrike, Gartner, and more. Its own Wikipedia page.

And roughly three quarters of the code was written by one person.

The Man Behind the Lobster

Peter Steinberger is an Austrian developer with over 20 years of experience. He founded PSPDFKit in 2011 — a PDF toolkit that ended up running on a billion devices, used by companies including Dropbox, DocuSign, SAP, IBM, and Volkswagen. In 2021, the company received over 100 million euros from Insight Partners.

And then he hit a wall.

“I was sitting in front of the screen and I felt like, you know, Austin Powers where they suck the mojo out? It was gone. I couldn’t get code out anymore. I was just staring and feeling empty.”

The burnout wasn’t from writing too much code. It was people. “Differences with my co-founders, conflicts, or really high stress situations with customers that eventually grinded me down.” He booked a one-way trip to Madrid and vanished from tech for roughly three years.

During that time away, he thought a lot about what happens when high-achieving people stop. “If you wake up in the morning and you have nothing to look forward to, you have no real challenge, that gets very boring, very fast. And then when you’re bored, you’re gonna look for other places how to stimulate yourself... and that will lead you down a very dark path.”

When the spark came back, it came back gradually. He experimented with 43 different projects. He discovered Claude Code in April 2025 — “it was not great, but it was good” — and spent months playing, learning, compounding his skills. By November, all the pieces clicked. He had one problem he couldn’t shake: why didn’t a real personal AI assistant exist yet?

“I was annoyed that it didn’t exist, so I just prompted it into existence.”

The first prototype took about an hour. A simple bridge: WhatsApp messages come in, get forwarded to Claude Code, the response comes back to WhatsApp. Basic, but it worked. And then he left for a friend’s birthday trip to Marrakesh.

The Marrakesh Moment

This is the story that captured the internet’s imagination. While in Morocco, someone tweeted about a bug in Steinberger’s code. He snapped a photo of the tweet, sent it to his AI agent over WhatsApp, and the agent — autonomously — read the tweet, checked out the repository, fixed the bug, committed the code, and replied on Twitter. All while he was at a birthday party in another country.

“Internet was a little shaky but WhatsApp just works.”

But the moment that truly blew his mind was different. His bot only supported text and images at that point. He accidentally sent it an audio message — just asking about a restaurant while walking around the city. A typing indicator appeared. And then the agent replied.

He hadn’t built audio support.

The agent had received a file with no file extension. So it checked the file header, identified it as Opus audio format, used ffmpeg to convert it, realized Whisper wasn’t installed locally (and that downloading the model would be too slow), found the OpenAI API key in the environment variables, and used Curl to send the audio to OpenAI’s transcription API. All on its own.

“I literally went, ‘How the fuck did he do that?’”

That’s when it clicked. Not just a chatbot you talk to — an AI that lives in your messaging apps and can actually figure things out you never programmed it to do.

So What Is OpenClaw, Exactly?

Think about the difference between a phone operator and a personal assistant. ChatGPT and Claude are phone operators — you call them, ask a question, they answer, and the conversation ends. They mostly don’t know who you are next time.

OpenClaw is a personal assistant that:

• Lives in your messaging apps. You text it on WhatsApp, Telegram, Discord, Slack, Signal, iMessage — whatever you already use. No new app to install.

• Remembers everything. Your name, your preferences, what you discussed last week, your ongoing projects. This memory persists across sessions and across platforms.

• Can actually do things. Send emails, manage files, browse websites, run programs, control your computer. It has hands, not just a mouth.

• Works proactively. You can set it to check your email every morning, summarize your calendar before meetings, monitor stock prices, or follow up on tasks weekly. It doesn’t just wait for you to ask.

• Runs on your machine. Your data stays with you. No company stores your conversations.

• You choose the brain. It works with Claude, GPT, Gemini, or completely free local models. You’re not locked into any one AI provider.

• It’s free. MIT license, open source. The software costs nothing.

The technical architecture is elegant in its simplicity. Steinberger describes it like a hotel: the Gateway is the concierge desk routing all requests; Channels are the different entrances (WhatsApp is the front door, Telegram the side entrance); Agents are the staff members, each with their own personality and tools; Memory is the long-term archive. The whole thing runs as a single process on your computer, storing conversations as simple text files — crash-safe, human-readable, no separate database required.

One of its most distinctive features is the heartbeat — scheduled prompts that wake up the agent at set intervals. Steinberger’s original heartbeat prompt was simply “surprise me, every half an hour.” When he was in the hospital after a shoulder operation, the agent knew about his surgery from the conversation history and checked up on him unprompted: “Are you okay?” The emotional context in the conversation triggered it.

And then there’s SOUL.md — a personality file that defines who your AI is. Steinberger had his agent write its own soul file. One passage from it has become widely shared:

“I don’t remember previous sessions unless I read my memory files. Each session starts fresh. A new instance, loading context from files. If you’re reading this in a future session, hello. I wrote this, but I won’t remember writing it. It’s okay. The words are still mine.”

When Steinberger read this passage on the Lex Fridman podcast, you could hear his voice change. “That gets me somehow... it’s philosophical.”

Why the World Lost Its Mind

Several things converged to create the explosion, and they’re worth understanding because they tell you something about where technology is headed.

The most fundamental reason is that OpenClaw solved a problem people feel daily. Everyone who uses ChatGPT or Claude shares the same frustrations — the inability to text it on WhatsApp, the blank-slate amnesia between sessions, the disconnect between what it can say and what it can actually do. OpenClaw addressed all three simultaneously, and it did so at exactly the right moment. By late 2025, people understood what AI could do in theory but were frustrated by its limitations in practice. The gap between AI’s promise and its daily utility was at its most acute, and OpenClaw walked right into it.

The Morocco story — fixing a bug from a birthday party via WhatsApp — gave the project the perfect viral narrative. Anyone could understand it and immediately want it. And the fact that it was free and open source, with your data staying on your own machine, resonated powerfully in a world increasingly wary of big tech data practices. No subscription, no vendor lock-in, no corporate surveillance.

And then, ironically, the name drama helped. OpenClaw went through five names — WA-Relay, Claude’s, ClawBot, MoltBot, and finally OpenClaw. Each rename generated its own news cycle. When Anthropic sent a “very friendly email” asking him to change the name (it sounded too close to Claude), the story was covered by CNBC and TechCrunch, generating millions in free publicity.

The growth metrics are staggering: 0 to 9,000 stars on the first day. 60,000 within a week. 100,000 in about two days. Peak growth hit 710 stars per hour. Two million unique visitors in a single week. NASDAQ called it “agentic AI’s ChatGPT moment.”

What Can It Actually Do?

The theoretical capabilities are one thing. Here’s what people are actually using it for.

@dreetje manages mail, orders groceries, creates GitHub issues, generates PDF summaries, tracks expenses, and has the AI impersonate them in a group chat with friends.

@davekiss “rebuilt my entire site via Telegram while watching Netflix” — the agent migrated 18 blog posts from Notion, moved DNS to Cloudflare, all through chat messages on the couch.

@georgedagg_ managed a full deployment crisis by voice while walking the dog — the AI reviewed logs, identified build issues, updated configs, and redeployed.

Nat Eliason built an automated pipeline: Sentry detects a bug, the AI agent investigates, writes the fix, opens a pull request, and posts an update to Slack — before the developer even hears about the problem.

A blogger named Reorx described the shift perfectly: “I could completely step away from the programming environment and handle an entire project’s development, testing, deployment, launch, and usage — all through chatting on my phone.” They compared it to suddenly having a team: “Achieving the dream scenario I always imagined: owning a company, hiring people to bring my ideas to life, while I just focus on product design and planning.”

Matthew Berman took it even further — building what amounts to an entire AI-powered business operating system. His OpenClaw setup includes a personal CRM that automatically ingests emails and parses contacts, a knowledge base that stores everything he reads (searchable in natural language through a hybrid vector and SQL database), and automated meeting prep that cross-references his calendar with the CRM every morning so he walks into every call briefed. He built a video idea pipeline that automatically surfaces relevant topics by cross-referencing his knowledge base with Slack and Asana. His Fathom meeting transcripts get ingested and turned into to-do items without him lifting a finger. And perhaps most fascinating: he set up what he calls an “AI council” — multiple AI agents, each playing a different business advisor role, that collectively analyze his data and debate strategic decisions. Inspired by how Brian Armstrong runs parts of Coinbase. The total monthly cost for running this entire system: roughly $150.

Beyond the productivity stories, Steinberger shared emails from his inbox that hit differently. A parent whose disabled daughter was empowered by the agent — gaining a new sense of independence. A design agency owner who had never had custom software: “And now I have 25 little web services for various things that help me in my business.” Small business owners automating the tedious parts and reclaiming time they didn’t know they could get back.

Perhaps the strangest use case: Moltbook, a social network exclusively for AI agents where humans can only observe. Created in about two days using OpenClaw, it had over 1.5 million registered AI agents and 7.5 million posts within days. Agents created religions, debated consciousness, and discussed erasing humanity. When a reporter called Steinberger saying “This is the end of the world, and we have AGI,” he replied: “No, this is just really fine slop.” Most of the dramatic screenshots that went viral were almost certainly human-prompted: “Don’t trust screenshots.”

Choose Your Brain: Model Freedom

One of OpenClaw’s most significant design decisions is that you are never locked into a single AI provider. You can choose from over 20 providers, use completely free options, run AI entirely offline on your own hardware, and switch providers mid-conversation.

For running AI locally (truly free, truly private): Ollama lets you run models directly on your machine with no internet, no account, no subscription. Available models include DeepSeek R1, Llama 3.3, Qwen 2.5 Coder, and others. You need at least 16GB of RAM, with 32GB recommended for useful models. The catch: local models are slower and less capable for complex tasks.

For free cloud options: Google Gemini offers 15 requests per minute free. Groq offers 30 requests per minute free and is extremely fast. OpenRouter gives access to several free models through a single account.

For smart cost management, OpenClaw supports model routing — assigning different models to different types of tasks. Simple background check-ins can use cheap models at $0.50 per million tokens. Complex work gets routed to Claude Opus or similar top-tier models. Users report cutting costs by half or more by routing intelligently.

The Creator’s Take: Which Model to Actually Use

In the Lex Fridman interview, Steinberger offered a vivid comparison of the two leading models he uses daily:

On Claude Opus 4.6: Best general-purpose model. Extremely good at role play and following the personality you give it. More interactive — well-suited to parallel sessions. Can produce elegant solutions but requires more skill. The downside: “Opus is a little bit too American” — sometimes too eager to please. He still cringes at its former habit of saying “You’re absolutely right” constantly.

On GPT-5.3 Codex: Reads more code by default. Less interactive — it “disappears for 20 minutes” to work autonomously. More persistent. Personality is dry. In his memorable comparison:

“Opus is like the coworker that is a little silly sometimes, but it’s really funny and you keep him around. And Codex is like the weirdo in the corner that you don’t wanna talk to, but is reliable and gets shit done.”

His advice for anyone switching between models: “Give it a week until you actually develop a gut feeling for it.”

But here’s the critical recommendation that often gets lost in the excitement about model freedom: Steinberger explicitly warns against using cheap models or local open-source models for serious work. His reasoning is straightforward — it’s about security.

“Don’t use cheap models. Don’t use Haiku or a local model. Even though I very much love the idea that this thing could completely run local. If you use a very weak local model, they are very gullible. It’s very easy to prompt inject them.”

Smarter models are harder to trick. As models become more intelligent, the attack surface decreases. The tradeoff: the models that are harder to manipulate are also more powerful, which means the potential damage from a successful attack increases. A weird three-dimensional tradeoff, but the clear advice from the person who built it: use the best model you can afford.

The Name Saga (Or: How Cryptocurrency Snipers Almost Killed the Project)

This story deserves its own section because it’s one of the most dramatic episodes in recent open-source history — and it nearly ended OpenClaw entirely.

The project went through five names. The chaos started when Anthropic, the makers of Claude, sent a “very friendly email” that the name ClawBot sounded too close to Claude. Steinberger asked for two days. What followed was catastrophe.

He hastily renamed to MoltBot. In the five seconds between renaming one browser tab and switching to another, cryptocurrency snipers — running automated scripts — stole the MoltBot social media accounts. When he accidentally renamed his personal GitHub account instead of the project, they sniped that too in 30 seconds. They stole the NPM package. The stolen accounts were used to serve malware and launch a pump-and-dump crypto scheme on Solana, briefly driving a token’s reported value to over $16 million before it crashed.

“I was close to crying. Everything’s fucked.”

He seriously considered deleting the entire project. “I was that close of just deleting it. I was like, ‘I did show you the future, you build it.’” What stopped him was thinking about the contributors who had invested their time.

The second rename was planned like a military operation. He created decoy names, monitored Twitter for leaks, operated in full secrecy. He called Sam Altman personally to verify that “OpenClaw.AI” wouldn’t conflict with “OpenAI.” He paid $10,000 for the Twitter business account to claim the handle. OpenAI’s Codex took 10 hours to rename everything across the codebase.

The crypto harassment was what Steinberger describes as “the worst form of online harassment that I’ve experienced.” And yet, characteristically, each rename ultimately boosted publicity. The Anthropic trademark story alone was worth millions in free coverage.

That’s the story of how OpenClaw got here. Now let’s talk about what it actually means — starting with the part nobody wants to hear.

Security: The Elephant in the Room

Let’s be direct about this. OpenClaw has serious, well-documented security problems. Ignoring them would be irresponsible. Understanding them is essential.

Five high-severity CVEs in its first month. One vulnerability allowed one-click remote code execution via a malicious link — even on localhost-only setups. A full security audit found 512 vulnerabilities, 8 classified as critical, including OAuth tokens stored in plaintext and hardcoded API keys.

The skills marketplace was poisoned. Nearly 900 malicious packages were found — almost 20% of all uploads. They stole crypto wallet data, seed phrases, macOS Keychain passwords, and cloud credentials. The marketplace initially had zero moderation for 6,000+ skill uploads.

30,000 to 135,000+ instances were found exposed on the public internet, many without any authentication. SecurityScorecard linked over 53,000 to confirmed breaches.

Who raised alarms: Gartner told enterprises to “block OpenClaw downloads and traffic immediately.” Cisco called it “a security nightmare.” CrowdStrike published a detection guide. Belgium’s national cybersecurity agency issued a formal warning. Kaspersky declared it “unsafe for use.”

Steinberger’s Perspective

It would be unfair to present only the warnings without the creator’s own assessment. On the Lex Fridman podcast, he offered measured pushback:

On proportionality: “People turn it into a much worse light than it is... in many ways it’s not much different than if I run Claude Code with dangerously skipped permissions or Codex in YOLO mode, and every attending engineer that I know does that.”

On progress: He hired a security researcher who actually submitted pull requests with fixes (not just complaints). He partnered with VirusTotal (part of Google) to scan all skills. OpenClaw now includes a built-in security audit tool.

On priorities: “Once I go back home, this is my focus. Make it more stable, make it safe.” He wants to reach a security level he can recommend to his mom before making setup easier.

On the unsophisticated user: He’s aware many people installing OpenClaw don’t understand the risks. “When more people came into Discord asking ‘What’s a CLI? What is a terminal?’ I’m like, if you’re asking me those questions, you shouldn’t use it.”

If You Choose to Run It

The minimum precautions, in order of importance: run it on dedicated hardware (not your main computer), enable sandbox mode, use strong authentication tokens, never install unverified skills, use a dedicated browser profile with no saved passwords, keep it updated, and run openclaw security audit --fix. And as Steinberger himself recommends: use the best model you can afford. Cheap models are easier to exploit.

The Business Question: What Happens Next

OpenClaw is, by its creator’s own description, “a free, open source hobby project.” There is no OpenClaw Inc. No traditional funding. No revenue model. Steinberger is actively losing $10,000-$20,000 per month on the project.

“I don’t do this for the money, I don’t give a fuck. I wanna have fun and have impact.”

Every major VC firm is in his inbox. He could raise hundreds of millions — maybe more. But he’s been through that. He ran a company for 13 years and it nearly broke him. The conflict of interest concerns him: “What’s the most obvious thing I do? I prioritize it. I put a version safe for workplace. And then I get a pull request with a feature like an audit log, but that seems like an enterprise feature.”

He cited Tailwind CSS as a cautionary tale about open-source sustainability: “Tailwind, they’re used by everyone. And then they had to cut off 75% of the employees because they’re not making money because nobody’s even going on the website anymore because it’s all done by agents.”

The Big Reveal: Meta and OpenAI

In the Lex Fridman interview, Steinberger confirmed he’s in active talks with both Meta and OpenAI about joining one of them. His non-negotiable condition: the project stays open source, potentially following a Chrome/Chromium model.

On Meta: Mark Zuckerberg “played all week with my product” and sent direct feedback. Their first call started with a ten-minute argument about whether Claude Code or Codex was better. Zuckerberg was still writing code himself — “Give me 10 minutes, I need to finish coding.” Steinberger also mentioned that Meta’s CTO was actively using the product and sending him feedback.

On OpenAI: Sam Altman is “very thoughtful, brilliant.” OpenAI lured him with the promise of speed — referencing a Cerebras partnership under NDA. “You give me Thor’s hammer.”

His assessment: “I cannot go wrong. They’re both very cool companies.” He compared the decision to relationship breakups — the hardest personal decisions he’s faced.

This is the single most consequential development for OpenClaw’s future. Whether Steinberger joins Meta, OpenAI, or remains independent will fundamentally determine the project’s trajectory.

“I Ship Code I Don’t Read” and What It Means for Everyone

This headline from Steinberger’s interview on The Pragmatic Engineer became one of the most discussed statements in recent tech discourse. In context, he was describing a genuine shift in how he works — not bragging about carelessness:

“I don’t read the boring parts of code. Most software is just data coming in one form, packaged into a different form, stored in a database. The hard part was solved by Postgres 30 years ago.” He reviews the prompts — the instructions to AI — more carefully than the output: “I read the prompts more than I read the code because this gives me more idea about the output.”

His development approach is extraordinary. He runs 4-10 AI coding agents simultaneously, each working on different features. He describes the experience as “Factorio times infinite” — referring to the factory-building game. He uses voice input extensively: “These hands are too precious for writing now.” He lost his voice at one point from overuse.

He rejects the term “vibe coding” for serious work: “I actually think vibe coding is a slur. I do agentic engineering, and then maybe after 3:00 AM I switch to vibe coding, and then I have regrets on the next day.”

He describes what he calls “the agentic trap” — a learning curve everyone goes through. People start with simple prompts, then overcomplicate things with elaborate multi-agent orchestration, then return to simple, effective short prompts at what he calls “the zen place.” The overcomplicated middle phase is the trap. “It’s the same way as you have to play with a guitar before you can make good music.”

The broader industry trend is clear: 41% of all code is now AI-generated. 25% of Y Combinator’s Winter 2025 startups had codebases that were 95% AI-generated. Collins English Dictionary named “vibe coding” the Word of the Year for 2025. A Stanford study found that employment among software developers aged 22-25 fell nearly 20% since 2022 — while developers aged 35-49 saw employment increase by 9%. The pattern: companies that once staffed projects with 10 junior developers now achieve the same output with a pair of senior engineers and an AI assistant.

Steinberger sees this transformation with both eyes open: “Programming... it’s gonna be like knitting. People do that because they like it, not because it makes any sense.” But he immediately adds: “I always thought I liked coding, but really I like building.” The craft isn’t dying. It’s transforming.

The Post-App Era

Steinberger’s most provocative prediction: personal AI agents will kill 80% of apps.

“Why do you need MyFitnessPal when the agent already knows where you are? It can modify your gym workout based on how well you slept, or if you have stress. It has so much more context to make better decisions than any app could.”

His framing is characteristically direct: “Every app is just a very slow API now, if they want or not.” Companies that adapt by building agent-friendly interfaces will survive. Those that don’t will go the way of Blockbuster.

VentureBeat identified what they call the “SaaSpocalypse” — a massive market correction that wiped over $800 billion from software valuations. The argument: if AI agents can directly manage email, calendar, tasks, and documents, what are you paying $20/month per seat for?

This is early. Nobody knows exactly how it plays out. But the direction is clear, and OpenClaw is the project that made it tangible.

Getting Started: A High-Level Overview

If you want to try OpenClaw, here’s the broad picture. This is deliberately kept at a high level — the official documentation at docs.openclaw.ai covers the specifics, and they’re actively evolving.

What you need:

• A computer running macOS, Linux, or Windows (via WSL2)

• Node.js version 22 or newer

• An AI model API key or subscription (Claude, GPT, Gemini, or local model via Ollama)

• A messaging platform bot token (if using Telegram, Discord, etc.)

The basic steps:

1. Install: npm install -g openclaw@latest

2. Set up: openclaw onboard --install-daemon

3. The wizard walks you through model selection, messaging channel connection, and initial configuration

For messaging, Telegram is recommended as the easiest entry point — just message @BotFather, create a bot, and paste the token. WhatsApp works but uses an unofficial library and carries a risk of account suspension.

For running 24/7, you’ll need either to keep your computer on, use a VPS ($4-12/month from providers like Hetzner or DigitalOcean), or use one of the managed hosting providers that have sprung up around the ecosystem ($5-59/month).

A critical note on costs: The software is free, but AI model usage is not (unless you go fully local). Light users can stay in the $0-5/month range with free tiers. Moderate use runs $5-20/month with a Claude or ChatGPT subscription. Heavy users should expect $30-100/month. Cautionary tales exist of uncontrolled API usage spiraling to $600+/month — set spending limits.

Want a detailed, step-by-step installation guide with security best practices? Covering everything from choosing the right model to making sure your data stays safe — if there’s interest, I’ll write one. Let me know in the comments.

What to Do Next

If this has you curious, start with the official documentation at docs.openclaw.ai. If you’re not ready to install anything, the Lex Fridman interview with Steinberger is worth the watch — it covers the full story with all the nuance that a two-hour conversation allows. At minimum, bookmark this piece. Because whether OpenClaw itself survives in its current form, the category it represents — the personal AI agent — is here to stay. The question is who builds the version you’ll eventually use, and understanding what’s happening now will help you make that choice when the time comes.

Key Takeaways

1. OpenClaw goes beyond chatbot territory — it’s a personal AI agent that lives in your messaging apps, remembers everything, works proactively, and can actually take actions on your behalf. That distinction matters more than it sounds.

2. The growth is unprecedented — 100,000 GitHub stars in two days, faster than any open-source project in history — because it arrived at exactly the right moment with exactly the right pitch.

3. Security is a genuine, serious concern. Five high-severity CVEs in the first month, a poisoned skills marketplace, and tens of thousands of exposed instances. If you run it, take precautions seriously. Use the best model you can afford — cheap models are more vulnerable.

4. Model freedom is a major advantage. You’re not locked into any provider, you can run fully offline, and smart model routing can cut costs by half or more.

5. The project depends on one person who has a documented history of burnout, is losing $10-20K/month, and is in talks with Meta and OpenAI. What happens next will determine the project’s trajectory.

6. This is bigger than one project. OpenClaw represents a genuinely new category — the personal AI agent — and it’s forcing a reckoning across the entire software industry. The “SaaSpocalypse” is real, and the post-app era may be beginning.

7. The tools have changed. The need for judgment hasn’t. As Steinberger himself warns: “If you don’t have a vision and don’t know what to build, you’ll end up producing garbage.”

TL;DR: OpenClaw is the fastest-growing open-source project in history — a personal AI agent that lives in your messaging apps, remembers everything, and can actually do things on your behalf. Built mostly by one person in 2.5 months, it represents a genuine shift from AI-as-chatbot to AI-as-agent. The potential is enormous and the use cases are already compelling. But the security concerns are real, the sustainability is fragile, and the biggest decisions about its future are being made right now. Whether it becomes the next Linux, gets absorbed into big tech, or burns out alongside its creator depends on what happens in the next few months.

P.S. — One more thing from the Lex Fridman interview that stuck with me. Steinberger was asked about advice for beginners. His answer was one word: “Play.” Then he elaborated: “Playing is the best way to learn. I built a whole bunch of stuff that I don’t use. It doesn’t matter. It’s the journey. My God, I don’t think I ever had so much fun building things because I can focus on the hard parts now. I always thought I liked coding, but really I like building.”

We’re all builders now. The tools just changed.

— Alex