OpenClaw

OpenClaw is an open-source AI assistant that runs on your own computer and talks to you in the chat apps you already use: Discord, iMessage, Slack, Teams, Telegram, WhatsApp and more. It is MIT-licensed with no paid tier. It is not mainly a coding agent. Checked 31 October 2026.

At a glance

What it is, and what it is not

This page sits in a list of coding agents, so the first thing to say is that OpenClaw is a different kind of tool. Claude Code, Cursor and OpenCode are built around a repository: you open a project, they edit files. OpenClaw is built around you. It is an assistant that stays on, lives on your hardware and answers in whatever messaging app you choose.

The README puts it this way: "State, memory, and credentials live on your hardware." The model and agent harness (Claude, Codex, local models) are plugins you can swap without changing anything else. So OpenClaw can drive coding tools. It can also do the rest of an assistant's work, which is why people compare it to coding agents anyway.

The shape of the system, from the README:

• The Gateway is the local control plane for sessions, tools, events and channel connections.
• The Control UI, CLI and terminal UI all connect to the Gateway.
• Channels bring the assistant to messaging services.
• Companion apps and nodes add voice, a canvas, camera, screen and device-local actions where the platform supports them.
• Tools, skills and plugins extend what it can do.

One Gateway runs either as a personal assistant on a laptop or as a shared team deployment; the README says configuration is the only difference.

Install

The installer supports macOS, Linux and Windows and provisions a supported Node.js runtime when needed.

# macOS, Linux, WSL2
curl -fsSL https://openclaw.ai/install.sh | bash

# Windows PowerShell
iwr -useb https://openclaw.ai/install.ps1 | iex

If you already manage Node.js, install the package. The README asks for Node 24.22.1 or newer, or 26.12.1 or newer, and recommends Node 26.

npm install -g openclaw@latest --allow-scripts=openclaw

That flag is for npm 12 or npm 11.16 and later. On npm 11.15 and earlier, leave out --allow-scripts=openclaw. The installation guide also covers Docker, Nix and other deployment paths.

First run

On a fresh install the installer starts onboarding for you. If you installed the package yourself, run:

openclaw onboard --install-daemon
openclaw gateway status
openclaw dashboard

Onboarding verifies model access, creates the workspace and configures the Gateway. The last command opens the Control UI, where you send a message to confirm the assistant works. Channel setup comes after that: WhatsApp, Telegram, Slack and the rest each have a page in the channels docs.

Models

OpenClaw works with hosted and local model providers. The model-providers page is an index that names:

You configure models in the Control UI under Settings, Models, or from the CLI:

openclaw models list
openclaw models set opencode/claude-opus-4-6

Custom and local servers go through the models.providers key with a custom base URL. One rule on that page saves confusion: "Adding provider auth does not change your primary model." Add auth first, then set the primary model explicitly.

Because no vendor sits behind OpenClaw, your cost is the provider's price, or your own hardware for a local model. See our OpenClaw vs Claude Code page for how that compares.

Skills, tools and plugins

OpenClaw has three extension routes: tools, skills and plugins. Skills use a format that will look familiar. A skill is a directory with a SKILL.md file: YAML frontmatter with a name and a description, then a markdown body.

---
name: release-notes
description: Draft release notes from a list of merged changes
---

When the user asks for release notes, group the changes by type and...

OpenClaw loads skills from several places, highest priority first: the workspace skills folder, <workspace>/.agents/skills, ~/.agents/skills, a managed state-directory folder, workshop skills, bundled skills and extra directories you configure. When two sources use the same skill name, the higher one wins.

ClawHub (clawhub.ai) is described as the public skills registry. You install with the openclaw skills command:

openclaw skills install @owner/<slug>
openclaw skills install git:owner/repo@ref
openclaw skills install ./path/to/skill --as my-tool
openclaw skills update --all
openclaw skills verify @owner/<slug>

Installs go to the active workspace by default, and --global installs to ~/.openclaw/skills. openclaw skills update only tracks ClawHub installs, so reinstall Git or local sources to refresh them. Skills can declare requirements (binaries, environment variables, config, operating system) in a metadata.openclaw block.

The .agents/skills path matches the one Antigravity uses, which makes it possible to share a skills folder between tools. See our skills page for the wider picture.

The skills page does not cover MCP, but the tools documentation does, and OpenClaw is an MCP client. See the MCP section below for how to add servers such as Playwright MCP and GitHub MCP.

Security and privacy

This is the section to read twice, because an always-on assistant that can run tools and receives messages from outside is a different risk than an agent you start yourself.

What the README says:

• Treat inbound messages as untrusted input. Direct-message channels pair unknown senders by default. You approve a pairing with openclaw pairing approve <channel> <code>.
• Tools run on the host for the main session unless you configure sandboxing. The README sends you to the security guide, the exposure runbook and the sandboxing guide before you connect other users or expose the Gateway remotely.
• Telemetry. By default the only thing OpenClaw sends itself is a daily version check. Anonymous feature statistics are opt-in. Setting update.checkOnStart: false disables both.
• Your prompts go to the model provider and chat platforms you configure, plus any diagnostics export you enable yourself.

So the privacy story is good in one respect and demanding in another. Nothing is stored by a vendor, because there is no vendor service. But you are the operator: you secure the Gateway, you choose the sandbox, and you decide who can message the assistant.

Adding MCP servers

OpenClaw is an MCP client, so the servers you already use elsewhere can be connected to it. The tools documentation gives three routes: the Control UI, the CLI and the config file.

In the Control UI, open Settings, MCP, choose Add server, name it, pick a transport (Streamable HTTP, SSE or stdio) and enter either the URL or the command and arguments. The docs say that writes an mcp.servers entry through the Gateway. Headers, environment values, OAuth settings, timeouts and tool filters live in a scoped config editor lower on the same page. From the CLI, openclaw mcp add takes a name and a command for a local stdio server.

One habit from the docs is worth copying: saving a definition proves nothing about whether the server works, so run the check:

openclaw mcp doctor <name> --probe
openclaw mcp status --verbose

The first command tests that the server answers. The second shows connection and authorization state, including sign-in instructions when OAuth has failed. Two other behaviours matter for an always-on assistant. When a server cannot start, the built-in client retries with exponential backoff, starting at 30 seconds and doubling up to 10 minutes. And when a server publishes usage instructions, OpenClaw keeps at most the first 2,000 characters per server in the prompt, to bound cost, so a server with long instructions may be partly cut.

The same docs describe the reverse direction. openclaw mcp serve runs OpenClaw as a stdio MCP server so another client can read and send channel conversations. That is powerful and risky in equal parts, since it hands a second program access to your messages. Enable it only for a client you trust, and read the security guide first. A separate plugin imports Claude Code and Claude Desktop instructions, MCP servers and skills into OpenClaw, which helps if you are moving over from them.

Scheduled work with automations

An assistant that stays on can act without being asked. OpenClaw's built-in scheduler, called automations, persists jobs, wakes the agent at the right time and delivers the output to a chat channel, a webhook or nowhere. The docs show a one-shot reminder as the quick start:

openclaw automations create "2027-02-01T16:00:00Z" \
  --name "Reminder" \
  --session main \
  --system-event "Reminder: check the automations docs draft" \
  --wake now \
  --delete-after-run

openclaw automations list
openclaw automations runs <job-id>

openclaw cron remains an alias for the same commands. Beyond reminders, the docs list cron schedules, inbound webhooks that let an outside service wake an agent, a Gmail trigger through Google Pub/Sub with a restricted reader, and an IMAP trigger that routes authenticated incoming mail to an isolated reader agent. The pattern in those last two is the one to notice: untrusted content, such as an email, goes to a restricted agent first, not straight to one that can run tools.

That design answers the main risk of scheduled assistants. A job that reads outside text and then acts is the shape of a prompt-injection attack. Start with jobs that only read and report, and keep acting jobs rare, narrow and logged. A useful first set is a morning summary of calendar and messages, a weekly reminder, and a webhook that posts a build result to a channel. Keep an eye on the run history for failures, because a job that silently stops is worse than one that was never set up.

A cautious first week

Because OpenClaw can run tools and reads messages, the order of setup matters more than the speed.

Treat adding a second user, a group chat or remote access to the Gateway as a separate project, with its own review of the exposure runbook. Each of those turns the question from "what can I make it do" into "what can anyone who reaches it make it do".

What it costs to run

OpenClaw itself has no price, so the cost is the model and the machine. A model route has three shapes. A hosted API bills per token, as on Claude API pricing. A subscription-style route, such as ChatGPT and Codex sign-in, uses a plan you already pay for, subject to that provider's terms. A local model costs electricity and hardware, and needs the memory sizing covered in How to self-host an LLM. An always-on assistant that handles dozens of short messages a day uses far fewer tokens than an agent running long coding sessions, but scheduled jobs and long contexts add up. Put a spend limit at the provider before you connect a channel, and look at the usage page after the first week.

For the machine, a small always-on computer or a low-cost VPS is enough for the Gateway when the model runs elsewhere. The cost rises if you run local models on the same box. Check the installation guide for the Docker and Nix paths if you prefer a container.

Personal or team: the two ways to deploy

The README says one Gateway serves both cases and that "configuration is the only difference." In practice the choice looks like this.

For a team, the README's own advice is blunt: read the security, exposure and sandboxing guides before you connect other users or expose the Gateway remotely. A pairing step for unknown senders is on by default for direct-message channels, which is a sensible starting point, but it is a starting point and not a full policy.

Questions people ask

Does it need the internet? Your model provider and chat platforms do, unless you use a local model and a channel that works locally. The README says OpenClaw itself only does a daily version check by default.

Can I run it without a chat app? Yes. The Control UI, CLI and terminal UI all connect to the Gateway, so a chat app is optional.

Who maintains it? The README names the OpenClaw Foundation as steward, described as an independent 501(c)(3). It also says there is no paid tier, hosted service or token.

What does the lobster mean? The README and banner use a lobster mascot. It has no bearing on the product, but it helps you recognize the project when you search.

Limits

• It is a general assistant, not a repository-first coding tool, so you will assemble the coding workflow yourself or hand work to a coding harness.
• You run and update it. There is no hosted tier to fall back on.
• Security depends on your configuration, as above.
• Documentation for each channel and provider is separate, so setup time grows with the number of integrations.
• The Kiro FAQ, as noted on our Kiro page, says using a Kiro subscription through third-party harnesses such as OpenClaw is not permitted. Check each provider's terms before routing a subscription through it.

Who it suits

• People who want a personal assistant in WhatsApp, Telegram or Slack that they control.
• Small teams that want one shared assistant and are comfortable running a server.
• Privacy-minded users who want to run local models and keep state on their own hardware.
• Developers who want to experiment with a hackable, MIT-licensed agent platform.

It does not suit someone who wants to open a repository and have an agent change it in ten minutes with no setup. That is what Claude Code and its peers do.

Moving to it from another agent

There is not really a migration, because the job differs. A better framing is "what should OpenClaw do that my coding agent does not":

1. Decide the channel. Start with one app you use daily.
2. Install and onboard on a machine that stays on, such as a home server or a small VPS.
3. Connect a model provider you already pay for, or a local model.
4. Keep coding work in your coding agent, and let OpenClaw handle reminders, messages and inbox tasks. If you want it to drive code, wire a coding harness in as a plugin.
5. Lock down pairing and sandboxing before inviting anyone.

Alternatives and duos

Our take

This is based on OpenClaw's README, which we read in full, and on the other pages named in the sources. We have not run it for this page.

OpenClaw is notable for how much it keeps out of someone else's hands: state, memory and credentials on your hardware, an MIT license, a foundation as steward and no paid tier. The scale of the project, with about 393,000 stars and a push the day before we checked, suggests a large community, though stars are a weak measure of quality.

The honest caveat is operations. A self-hosted assistant that reads your messages and can run tools is only as safe as its configuration. Treat the security guide as required reading, and start with one channel and the sandbox on.

If you want a coding agent, pick one from the list. If you want an assistant that lives in your chats and is yours, this is one of the most developed open options we know of.

Checklist

• Choose a machine that can stay on
• Install with the script and finish onboarding
• Run openclaw gateway status and send a test message
• Connect one channel and keep pairing approval on
• Configure sandboxing before connecting others
• Set update.checkOnStart: false if you do not want the version check