What is OpenClaw? How the self-hosted AI agent works (2026)
OpenClaw explained: the Gateway, the agent loop, memory, heartbeat, skills and chat channels, with a diagram that follows one message through the system.
OpenClaw is a free, open-source AI assistant that you run on your own computer or server and talk to through chat apps you already use, such as WhatsApp, Telegram, Discord, Slack or iMessage. It doesn’t come with an AI model: it connects to one you choose, from an API like Claude or OpenAI’s models or a local model, and gives that model tools (a browser, a shell, your files, scheduled jobs) so it can do things for you, not just answer.
We haven’t run OpenClaw ourselves. This explainer is built from OpenClaw’s own documentation and repository as of October 20, 2026; see how we work.
OpenClaw at a glance
| What it is | A self-hosted assistant: one “Gateway” process that links your chat apps to an AI model and its tools (docs) |
| Who makes it | The OpenClaw Foundation, a US 501(c)(3) non-profit announced in July 2026; created by Peter Steinberger (foundation post) |
| License and price | MIT, free. “No subscription. No hosted tier.” (openclaw.ai) You pay for the model and the machine it runs on |
| Latest stable release | v2026.10.5, published October 18, 2026 (GitHub) |
| Popularity | 392,933 GitHub stars and 82,541 forks on October 20, 2026 (GitHub) |
| Runs on | macOS, Linux and Windows (native or WSL2); needs Node 24.16+ or 26.1+ (install docs) |
| Talks through | More than 30 chat integrations on its Channels page, from WhatsApp and Telegram to Signal, Microsoft Teams and SMS |
The name has changed several times. The project began as a weekend hack called “WhatsApp Relay”, then became “Clawd”, a pun on Claude; its GitHub releases went out as warelay, clawdis and, from January 5, 2026, Clawdbot. After “Anthropic’s legal team politely asked us to reconsider”, it was briefly called Moltbot, and OpenClaw was announced on January 29, 2026 (OpenClaw blog, GitHub releases). The history in data covers what happened next.
How OpenClaw works: the parts
Everything runs through one long-lived process, the Gateway. It owns every messaging connection, and the apps that control OpenClaw (the command line, the web dashboard, the Mac app) connect to it over a WebSocket. There is one Gateway per machine, and by default it listens only on the machine itself, at 127.0.0.1:18789, with a token required even there (architecture, Gateway docs).
Around the Gateway sit the pieces you configure:
- Channels connect chat apps. WebChat is built in, Telegram ships bundled, and WhatsApp, Discord, Slack, iMessage, Signal and most others are official plugins you install when needed (Channels).
- Agents are personas, each with its own workspace, settings and conversation history. One Gateway can run several, and “bindings” decide which agent answers which chat (multi-agent routing).
- The model is whatever you connect: a provider API key, an existing Claude Code or Codex login, or a local model through Ollama or LM Studio (getting started, onboarding).
- Tools are the actions the model can take: run commands, read and edit files, search and fetch the web, drive a browser, send messages, schedule jobs and generate images or speech (tools).
- Skills and plugins extend it. A skill is a
SKILL.mdfile of instructions loaded into the prompt; a plugin adds code: new tools, channels or model providers (tools, skills and plugins).
Click any part below to see what it does, or step through what happens to one message.
Follow one message through OpenClaw
Each box is a part named in OpenClaw's docs. Use Next to follow a request from your phone to the reply, then the heartbeat that runs without you.
What happens to one message, in short
A message goes from your chat app to the Gateway, to an agent, which builds a prompt from its workspace files, skills and history, sends it to the model, runs whatever tools the model asks for, and replies. The agent loop docs describe the same six stages: intake, context assembly, model inference, tool execution, streaming and persistence. Two details matter for everything else in this series:
- Every model call carries a lot of context. Instructions, persona, memory, skills and tool definitions go in with each call, and a request that uses tools makes several calls. That drives the bill; see what OpenClaw costs to run.
- Tools run with your machine’s permissions. For the main session, commands run on the host unless you configure a sandbox, which is off by default (sandboxing). That drives the risk; see is OpenClaw safe?.
Memory: plain files, not a black box
OpenClaw keeps its memory as Markdown files in the agent’s workspace (the docs say “there is no hidden state”). MEMORY.md holds curated long-term notes and is loaded only in your main, private session; memory/YYYY-MM-DD.md is a daily log (memory).
The agent searches these files with a memory_search tool, by keyword, by meaning (vector search) or both, using a built-in SQLite engine by default. Semantic search needs an embedding model: OpenAI’s by default, with Gemini, Mistral, Voyage and local options such as Ollama or LM Studio as alternatives.
Two automatic jobs keep memory current. Before a long conversation is compacted, a silent turn writes the important parts to the daily note. A background “dreaming” job, on by default, later promotes the items worth keeping into MEMORY.md. Because it’s all plain text, you can read and edit what the assistant remembers. The docs also suggest backing the workspace up to a private Git repository, without secrets (workspace).
The heartbeat: why it acts without being asked
Most chat assistants only answer. OpenClaw also runs a heartbeat: a periodic turn in your main session, every 30 minutes by default (every hour if you sign in with Anthropic OAuth), in which the model can raise anything that needs your attention. When there’s nothing, it replies NO_REPLY and you see nothing (heartbeat). You can limit it to active hours or turn it off with 0m.
Automations are the scheduler behind it. They run one-off, interval or cron-style jobs (“summarize my unread email at 7am and send it on WhatsApp”) and deliver the result to a chat, a webhook or nowhere (automations). The Gateway has to be running for any of this to fire, which is the main reason people keep OpenClaw on an always-on machine; see where to run OpenClaw.
A heartbeat is a model call like any other, so it costs tokens even when it finds nothing. At the default 30 minutes, that’s 48 calls a day if it runs around the clock, each carrying your main session’s context.
Skills, plugins and MCP
Skills are folders with a SKILL.md file that teaches the agent a task or an API. They follow the AgentSkills format and load into the prompt when relevant. Plugins are code that adds tools, model providers, chat channels or hooks (skills, plugins).
Both come from ClawHub, “the public registry for OpenClaw skills and plugins”, which scans releases automatically (ClawHub). The docs are blunt about the risk: treat third-party skills as “untrusted code” and plugin installs “like running code”. Our guide to OpenClaw skills covers what’s there and how to vet one.
OpenClaw also speaks MCP, the Model Context Protocol. It can use tools from external MCP servers (still subject to your tool policy) and can act as an MCP server itself, exposing its conversations to other MCP clients (MCP).
What people use OpenClaw for
The project’s showcase lists about 85 builds that users posted publicly. These are their accounts, not results we checked, but they show the range. Filter by kind of task:
What people have built with OpenClaw
Each row is a build posted to OpenClaw's public showcase, described in our words.
| What it does | How |
|---|---|
| Morning email digest | A daily scheduled job summarizes unread email, sends it on WhatsApp and creates to-dos |
| One chat for mail, messages, calls and a password vault | One assistant handles email, reminders, GitHub issues, voice calls and a 1Password vault |
| Personal "operating system" | Daily briefings, weekly reviews, meeting briefs, school deadlines and invoices |
| Flight check-in | Finds flight details in email, then checks in and picks seats |
| Car purchase negotiation | Negotiated with dealers by browser, email and iMessage |
| Family meal planning | Yearly menu template, shopping lists sorted by aisle, weather-aware dinners and reminders |
| Smart home from one chat | Email, Home Assistant, a home lab, to-dos and shopping lists from one Telegram chat |
| Production fix by voice | Diagnosed failed builds, fixed configs, redeployed and opened a pull request while walking the dog |
| Coding agents overnight | Keeps notes, builds a work list and launches Codex agents that open pull requests |
| Website rebuilt from Telegram | Moved a site from Notion to Astro and its DNS to Cloudflare without a laptop |
| Agent fleet across 3 machines | 15+ agents clear email, review decks, build tools and coordinate coding workers through Discord |
Showing all 11.
Source: openclaw.ai/showcase, read October 20, 2026.
The pattern is clear: OpenClaw is most useful for tasks that span several apps (email, a website, a calendar) and that you’d otherwise do by hand at odd hours. Most of these examples also need it to hold real logins, which is why setup and security matter more than with a chatbot.
What it takes to run
- The software installs with one command on macOS, Linux or Windows (
curl -fsSL https://openclaw.ai/install.sh | bash, or a PowerShell equivalent), or through npm, Docker, Nix and others.openclaw onboardthen walks you through the model, the Gateway and your first channels; the docs put setup at “about 5 minutes” (install, getting started). - A machine that stays on, if you want it to answer and run jobs while your laptop sleeps. The docs give no general RAM figure; the Raspberry Pi page sets a floor of 1 GB of RAM and 500 MB of disk, and building the Docker image from source needs 6 GB of RAM (Raspberry Pi, Docker). Pre-built images skip that build. Compare a VPS, a Mac mini and a Pi in where to run OpenClaw.
- A model. No GPU is needed if you use an API model. Running the model yourself is possible but demanding; see OpenClaw with local models and the best model for OpenClaw.
- Upkeep. OpenClaw moves fast: GitHub shows five stable releases between October 10 and October 18, 2026 alone.
openclaw updatehandles upgrades, but its rollback doesn’t undo database changes, so the docs advise a verified backup before big updates (updating).
What OpenClaw is not
- Not a hosted service. There’s no official cloud version; the foundation says it has “no paid tier, no hosted service” (FAQ). Third parties offer one-click setups, which are covered in where to run OpenClaw.
- Not a model. Its quality depends on the model you connect.
- Not a multi-user security boundary. The docs describe “one trust boundary per gateway” and say it “is not a hostile multi-tenant security boundary”. People who shouldn’t share access need separate Gateways (security).
- Not the only option. Hermes Agent does the same job with stricter defaults, and several smaller projects copy the idea; see OpenClaw vs Hermes Agent and OpenClaw alternatives.
Is OpenClaw still relevant?
OpenClaw was the breakout open-source project of early 2026, and public interest has since cooled: US Google searches for “openclaw” went from about 1,000,000 a month in February and March to about 74,000 in August and September, according to DataForSEO’s keyword data. Development hasn’t slowed, though. The project moved to a non-profit foundation in July, ships several releases a week, and in September announced a free, self-hosted OpenClaw Enterprise for managing agents in sensitive environments, built with Red Hat and NVIDIA and still in early pilots (announcement).
So the question isn’t whether OpenClaw works. It’s whether it suits you better than the alternatives, at a price and risk level you accept. The rest of this series is built to answer that.
FAQ
What does OpenClaw do?
It runs an AI assistant on your own machine that you message from chat apps. With the tools you allow, it can browse, run commands, edit files, handle email and calendar tasks, and run scheduled jobs and check-ins on its own.
Is OpenClaw free?
Yes. The software is MIT-licensed and free, with no paid tier. You pay for the model (an API, a subscription login or your own hardware) and for the computer or server it runs on; see what it costs.
Who owns OpenClaw?
The OpenClaw Foundation, a US non-profit announced in July 2026, with Dave Morin as chair. Its creator, Peter Steinberger, has joined OpenAI and still stewards the project. The foundation says donors, OpenAI among them, don’t own or direct it (FAQ).
Is OpenClaw the same as Clawdbot or Moltbot?
Yes. It’s the same project: Clawd (released as Clawdbot) until late January 2026, Moltbot for a few days after that, and OpenClaw since January 29, 2026.
Does OpenClaw need a GPU?
No, not with an API model: the Gateway itself is a Node.js app that runs on a small server, a Mac mini or a Raspberry Pi. You need a GPU only to run the model locally.