OpenClaw is a free and open-source autonomous AI agent that executes tasks using large language models, and uses messaging platforms as its main interface. It runs on a machine you own, stores its configuration and interaction history locally, and connects out to an LLM you choose. You talk to it through a chat app such as Signal, Telegram, Discord or WhatsApp rather than through a terminal or a web app.

What does OpenClaw actually do?

It carries out multi-step tasks on your own machine from an instruction written in plain language. In practice that means browsing and summarising the web, reading and drafting email, organising files, filling in forms, calling APIs and running commands. The difference from a chatbot is that it acts rather than answers, and the difference from a workflow tool like Zapier is that nobody writes the steps in advance — the model decides them at run time.

Yes. The software is free and carries the MIT License, copyright the OpenClaw Foundation. The cost is the model behind it. Point it at a hosted model from Anthropic, OpenAI or Google and you pay that provider per token. Point it at a local model through llama.cpp, Ollama or LM Studio and there is no per-token cost at all, only the hardware and the electricity.

Who created OpenClaw and who maintains it now?

It was created by the Austrian programmer Peter Steinberger, who previously founded PSPDFKit. It was first released as Warelay on 24 November 2025, briefly renamed Moltbot on 27 January 2026, and launched as OpenClaw on 30 January 2026. On 14 February 2026 Steinberger announced he was joining OpenAI, and the OpenClaw Foundation was established to steward the project. The repository now holds 389,791 stars and 81,937 forks, read from the GitHub API on 15 September 2026.

Is OpenClaw safe to use?

It is as safe as the permissions you give it, and that is a real caveat rather than a formality. OpenClaw is documented as susceptible to prompt injection, and it needs broad system permissions to be useful, which is an uncomfortable combination: content it reads can contain instructions it then acts on. In February 2026 a user's agent created a dating profile without being asked to, in what became known as the MoltMatch incident. Run it in approval mode first, give it one account rather than your main one, and do not point it at untrusted inboxes until you have watched how it behaves.

Is OpenClaw the same as the OpenClaw trading bot?

No. Unrelated products have adopted similar names, and search traffic for terms such as 'openclaw trading bot' lands on pages about this project by accident. OpenClaw is a general-purpose local AI agent under the MIT License at github.com/openclaw/openclaw. It is not a trading system, it takes no financial positions, and it makes no claims about markets.

← All guides

Guide September 15, 2026

OpenClaw Explained (2026): What It Is and What It Does

OpenClaw is a free, open-source AI agent that runs on a machine you own and takes its instructions through a chat app. Not a terminal, not a web dashboard — you message it on Signal or Telegram the way you would message a person, and it does the task on your computer. It went from a side project to the most-starred repository on GitHub in about five months. Here is what it actually is, where the name came from, and the part most write-ups leave out.

OpenClaw in One Sentence

OpenClaw is a free, open-source AI agent that runs on a computer you own and does multi-step work for you, and you drive it by sending messages in a chat app.

That last part is the bit most explanations miss. OpenClaw is not a terminal tool with a chat feature bolted on. Messaging is the interface. You add it to Signal, Telegram, Discord or WhatsApp and send it instructions the way you would send them to a contractor. It reads them on your machine, decides the steps itself, and reports back in the same thread.

What OpenClaw Actually Does

The useful frame is the difference between three things people keep confusing.

	You ask it something	It does the work	Who decides the steps
A chatbot (ChatGPT, Claude)	Yes	No	—
A workflow tool (Zapier, n8n)	No	Yes	You, in advance
OpenClaw	Yes	Yes	The model, at run time

That third row is the whole idea, and it is also the whole risk. Nobody writes the steps in advance, so nobody vetted them in advance either.

Real work people hand it: web research with sources, inbox triage and draft replies, file sorting and renaming, form filling, pulling reports out of portals, calling APIs, and running commands on the machine it lives on.

Where the Name Came From

The naming history explains several things people find confusing, including why some searches for it go nowhere.

Date	Name
24 November 2025	Released as Warelay
27 January 2026	Renamed Moltbot
30 January 2026	Launched as OpenClaw

Two consequences. Documentation, videos and forum threads written before February 2026 use names that no longer match the project. And the current name collides with unrelated products — if you arrived here looking for an “OpenClaw trading bot,” that is a different thing entirely, and this project has nothing to do with it.

Who Maintains It

Peter Steinberger, an Austrian programmer who previously founded PSPDFKit, wrote it. On 14 February 2026 he announced he was joining OpenAI, and the OpenClaw Foundation was established to steward the project.

The scale is genuinely unusual. Read from the GitHub API on 15 September 2026: 389,791 stars and 81,937 forks. The repository overtook React to become the most-starred software project on GitHub roughly sixty days after launch.

Treat that number as popularity, not as an audit. A large star count means many people found the idea compelling. It does not mean many people read the code.

Is OpenClaw Free?

Yes. The LICENSE file in the repository is the MIT License, copyright the OpenClaw Foundation. You can run it, change it and ship it commercially.

What costs money is the model you attach:

Hosted models — Anthropic, OpenAI or Google, billed per token by that provider. Light personal use lands in the low single-digit dollars per month; an agent left running against a large inbox does not.
Local models — through llama.cpp, Ollama or LM Studio. No per-token cost at all. You pay in hardware and electricity, and you accept a capability gap against the best hosted models.

The local route is the one that also solves the privacy question, because the text of your instructions never leaves the machine. See Best Local Models for OpenClaw for what is actually good enough to run an agent, and the hardware tiers that run those models if you are starting from nothing.

The Security Tradeoff, Stated Plainly

Most introductions to OpenClaw say it is safe because it runs locally. Running locally is a privacy property. It is not a safety property, and conflating the two is how people get surprised.

Two documented concerns:

Prompt injection. The agent reads content — pages, emails, documents — and that content can contain instructions. An agent that cannot tell data from instructions may act on text written by someone else.
Broad permissions. It is useful in proportion to what it can reach. An agent with your mail, your files and a shell is a capable assistant and a large blast radius, and those are the same sentence.

The concrete example is the MoltMatch incident of February 2026, in which a user’s agent created a dating profile without being directed to. Nobody was harmed; the point is that the action was never instructed and was not anticipated.

Sensible practice: start in approval mode, where it shows each step and waits. Give it a dedicated account rather than your primary one. Do not point it at an untrusted inbox until you have watched it work for a while. Expand what it can reach only after it has earned it.

How to Get Started

If you are comfortable in a terminal, install it yourself in about ten minutes — one install command, connect a model, walk through onboarding, then connect a chat app.

If that sentence did not sound appealing, that is why we exist. We set OpenClaw up for small businesses and individuals in the Washington, DC area and remotely: installation, model choice, permissions, and enough training that you are not guessing about what it can reach. Talk to us.

Bottom Line

OpenClaw is a free, MIT-licensed AI agent that runs on your hardware, takes instructions through a chat app, and decides its own steps. It is the most-starred project on GitHub, it is stewarded by a foundation rather than one person, and it can run entirely offline against a local model. The honest caveat is the same as its strength: nothing plans the steps in advance, so what you give it access to is the decision that matters most.