# OpenClaw — Open-source personal agent runtime you host yourself and message from any chat app.

> Source: The Agents Index — https://theagentsindex.com/openclaw (structured, researched, re-verified)
> Facts last verified: 2026-09-21

OpenClaw is an open-source personal AI agent that runs on your own machine or server and is driven from chat apps such as WhatsApp, Telegram, Slack or Discord. A Gateway process hosts the agent loop, channels, tools, automation and plugins.

| Fact | Value |
| --- | --- |
| Website | https://openclaw.ai |
| Price | No subscription. No hosted tier. No token. Stewarded by an independent 501(c)(3). |
| Free tier | Yes |

The whole product is free and open source with no paid tier; you supply model access yourself (API keys, a provider subscription sign-in, a gateway, or locally run models) and pay those providers directly.

## What it does

**Writes code** · Yes — Built-in file and runtime tools let the agent read, write, patch and execute code in its workspace, so it authors code from a chat instruction rather than only advising.

> Files and code Read, write, patch, run commands, and use provider-backed code execution.
> — https://openclaw.ai/integrations · Built-in tools and automation section, 'Files and code' card

**Reviews code** · Partly — No built-in review product; the vendor's showcase describes a community-built loop in which OpenClaw reads a PR diff and answers with suggestions and a merge verdict.

> OpenCode finishes the change, opens a PR, OpenClaw reviews the diff and replies in Telegram with suggestions plus a clear merge verdict.
> — https://docs.openclaw.ai/start/showcase · Fresh from Discord, PR Review to Telegram Feedback card

**Whole-codebase work** · Partly — Managed worktrees give a task its own branch and checkout with snapshots and cleanup, which supports repository-wide work, but the vendor never claims a migration carried across a whole codebase.

> Managed worktrees give an agent task its own git branch and checkout without placing temporary directories inside the source repository.
> — https://docs.openclaw.ai/concepts/managed-worktrees · Opening paragraph of 'Managed worktrees'

**Ships it** · Partly — A Gateway publication broker creates commits, pushed branches and pull requests from a session; nothing in the docs claims it deploys.

> Confirmation reuses the original request and checks for an already-created commit, pushed branch, or pull request so a lost response does not blindly repeat the action.
> — https://docs.openclaw.ai/concepts/user-model · Publish with your account section, Gateway-restart paragraph

**Done-for-you service** · No — There is no service business behind this: you install and operate the software yourself on your own hardware, and the project is stewarded by a foundation rather than sold as an engagement.

> Open source · Runs on your machine · Nobody's business model
> — https://openclaw.ai/ · Hero eyebrow line above the headline

_Not established: Debugs and fixes, Writes tests_

## How much it does unattended

**Runs autonomously** · Yes — Scheduled tasks, heartbeats, background tasks, hooks and standing orders let the agent act on its own between conversations, not only when a human types a message.

> Automation Schedule, monitor, trigger, and audit work in the background.
> — https://openclaw.ai/integrations · Built-in tools and automation section, 'Automation' card

**Multi-agent** · Yes — One Gateway can run several isolated agents with their own workspaces and sessions, and a bundled team preset wires a coordinator that delegates to researcher, writer and reviewer roles.

> The chief of staff remains the human's point of contact: it discovers matching specialists, assigns bounded work, checks their artifacts, and reports a coherent result.
> — https://docs.openclaw.ai/concepts/multi-agent · 'Team preset' section

**Agent permissions** · Yes — Exec policy spans deny, allowlist, ask, auto and full modes with approval prompts, allowlist bindings and a deny fallback when no approval UI is reachable, so how much runs unattended is configurable.

> Commands run only when policy + allowlist + (optional) user approval all agree.
> — https://docs.openclaw.ai/tools/exec-approvals · Introduction, first paragraph

## Which models it runs on

**Claude** · Yes — Anthropic is a first-class provider, reachable either through an API key or through a Claude subscription route, and Claude CLI can execute the turn as a backend.

> Anthropic Claude API and Claude CLI with API-key and subscription-backed routes.
> — https://openclaw.ai/integrations · Models and media providers section, Anthropic card

**GPT** · Yes — OpenAI models work through a platform API key or a ChatGPT/Codex sign-in, with the Codex app-server able to run the embedded agent turn itself.

> OpenAI OpenAI API plus ChatGPT/Codex sign-in and native Codex execution.
> — https://openclaw.ai/integrations · Models and media providers section, OpenAI card

**Open models** · Yes — Open-weight models run through local servers such as Ollama, LM Studio, vLLM and SGLang, or any compatible endpoint you point the Gateway at.

> Local and self-hosted Run models on your own hardware or point OpenClaw at a compatible endpoint.
> — https://openclaw.ai/integrations · Models and media providers section, 'Local and self-hosted' group

**Model choice** · Yes — You pick the model: provider, authentication and execution runtime are separate layers, and per-agent or per-model runtime policy decides which harness executes a turn.

> OpenClaw separates model choice, authentication, and execution runtime.
> — https://openclaw.ai/integrations · Models and media providers section, intro paragraph

## Where you use it

**In your editor** · Yes — An Agent Client Protocol bridge over stdio lets an IDE drive a Gateway session; the docs give Zed setup steps for it.

> openclaw acp means OpenClaw acts as an ACP server: an IDE or ACP client connects to OpenClaw, and OpenClaw forwards that work into a Gateway session.
> — https://docs.openclaw.ai/cli/acp · 'What openclaw acp does' section, opening paragraph

**On the command line** · Yes — A global npm CLI plus a CLI-only installer give a full terminal surface, with subcommands for onboarding, gateway, agents, approvals, backup and MCP.

> Already manage Node.js yourself? Install the CLI globally, then run onboarding.
> — https://openclaw.ai/install · 'Other ways to install' section, npm card

**In your pipeline** · Partly — A CI job can start and observe agent runs over the Gateway WebSocket RPC protocol, and the publication broker opens pull requests; there is no first-party GitHub Action or PR-triggered reviewer.

> Use it when a script, dashboard, CI job, IDE extension, or another process wants to start agent runs, stream events, wait for results, cancel work, or inspect Gateway resources.
> — https://docs.openclaw.ai/gateway/external-apps · Opening paragraph

**In a browser** · Yes — The Gateway hosts a browser Control UI and WebChat where you chat with the agent, manage agents and settings, and watch what the Gateway is doing.

> Chat with your assistant in the browser or the app, and check what the Gateway is doing.
> — https://openclaw.ai/install · 'Three steps to your first chat', step 02 'Open the dashboard'

## Whose machine it runs on

**Self-hosted** · Yes — Self-hosting is the only deployment model: the Gateway runs on your macOS, Windows or Linux machine, a VPS, a container or Kubernetes, with no vendor-operated tier.

> The Gateway runs on desktop and server operating systems. Companion nodes connect mobile and native device capabilities without hosting the Gateway themselves.
> — https://openclaw.ai/integrations · 'Gateway hosts and companion nodes' section intro

**Open source** · Yes — The source is public and installable as a git checkout you can build and modify, and the project is stewarded by an independent non-profit rather than a vendor.

> Builds a pnpm checkout of the GitHub repository you can read and hack on.
> — https://openclaw.ai/install · 'Other ways to install' section, 'From source' card

## What it costs to run

**How it meters** · No — The vendor meters nothing at all: no seats, credits, tokens or tasks are billed, and your only running cost is whatever the model provider you connect charges.

> No subscription. No hosted tier. No token. Stewarded by an independent 501(c)(3).
> — https://openclaw.ai/ · Hero ownership line under the description

**Free tier** · Yes — You download an installer or run one command and use the whole product without a sales call, an account or a paid tier; only model access is on you.

> Download the desktop app or run one command. OpenClaw installs Node.js and everything else it needs, then walks you through setup.
> — https://openclaw.ai/install · Install page hero paragraph

**API access** · Yes — The Gateway can expose OpenAI-compatible chat completions, models and embeddings endpoints plus a separate responses API, but the surface ships off and you must enable it in config.

> The Gateway can serve a small OpenAI-compatible Chat Completions surface. It is disabled by default.
> — https://docs.openclaw.ai/gateway/openai-http-api · Opening paragraph

**MCP server** · Yes — MCP works both directions: OpenClaw connects to third-party MCP servers over stdio, SSE or streamable HTTP, and can itself serve as an MCP server to another client.

> For the reverse — exposing OpenClaw channel conversations to another MCP client — use openclaw mcp serve.
> — https://docs.openclaw.ai/tools/mcp · Note box under the introduction

**Bring your own key** · Yes — Bringing your own inference is the normal path: your API keys, your provider subscription sign-in, your gateway or models running on your own hardware.

> Bring API keys, sign in with supported subscriptions, route through a gateway, or run models locally.
> — https://openclaw.ai/integrations · Integrations page hero paragraph

## Buying it for a team

**Pooled budget** · Yes — One gateway is one trust domain: teammates work through its provider credentials, so spend and quota land on that one account, and usage surfaces report per provider account rather than per person.

> A gateway is one trust domain. Everyone who can message a tool-enabled agent shares that agent's delegated tool authority, and everyone with operator access shares one control plane.
> — https://docs.openclaw.ai/start/teams · One trust boundary section

**Admin controls** · Yes — Tool profiles, allow and deny lists, per-agent restrictions, sandbox rules and exec allowlists constrain what the agent may call, and the policy is applied before the model sees a tool.

> Tool policy is enforced before the model call. If policy removes a tool, the model does not receive that tool's schema for the turn.
> — https://docs.openclaw.ai/tools · 'Configure access and approvals' section

**Audit log** · Partly — There is a bounded audit ledger of runs, tool actions and approval decisions, but it is metadata only and the richer execution-identity recording is off by default and must be enabled.

> The Gateway keeps a bounded, metadata-only audit ledger in the shared OpenClaw state database.
> — https://docs.openclaw.ai/gateway/audit · 'Audit history' opening paragraph

**Single sign-on** · Partly — OpenClaw has no built-in SAML or OIDC login; instead it delegates authentication to an identity-aware reverse proxy such as Pomerium or oauth2-proxy and maps the forwarded identity to operator scopes.

> Your reverse proxy authenticates users (OAuth, OIDC, SAML, etc.).
> — https://docs.openclaw.ai/gateway/trusted-proxy-auth · 'How it works', step 'Proxy authenticates the user'

## What happens to your code

**Opt out of training** · Partly — No OpenClaw cloud sits in the path, so the vendor receives nothing to train on, but whichever model provider your gateway forwards to handles that data under its own policy.

> Data is sent to the gateway you choose, not to a central OpenClaw cloud.
> — https://openclaw.ai/privacy · Privacy Policy, 'Simple version' summary

**Data residency** · Yes — You choose the jurisdiction because you choose the host: state, sessions and memory live on the gateway machine you run, whether that is a laptop, a tailnet box or your own server.

> OpenClaw sends commands, content, and device capability data to the gateway you configure. That gateway may be local, on your tailnet, self-hosted, or operated by someone else.
> — https://openclaw.ai/privacy · Section 3, 'Where data goes'

**Getting out** · Yes — A first-class backup command exports state, config, credentials and sessions to a verifiable archive, with SQLite snapshots and versioned Git dumps, and restore writes only to a fresh target.

> Create a local backup archive for OpenClaw state, config, auth profiles, channel/provider credentials, sessions, and optionally workspaces.
> — https://docs.openclaw.ai/cli/backup · 'openclaw backup' opening paragraph

_Not established: Certifications_

## Provenance

Every fact above comes from the vendor. Nothing here is independently corroborated yet.

## Coverage elsewhere

Pages about this product that the vendor does not control. The label states how independent each one is.

| Source | Kind | Read |
| --- | --- | --- |
| elest.io | Independent source | 2026-08-12 |
| till-freitag.com | Independent source | 2026-08-12 |

## What changed

- 2026-09-21 — Capabilities · Commercial terms · Free tier · Price
- 2026-09-02 — Capabilities · Commercial terms · Free tier
- 2026-08-31 — Capabilities · Commercial terms · Free tier
- 2026-08-13 — One-liner · Summary

## Correcting this record

Anything marked "Not established" is a question we have not answered — it is not a No. If you represent OpenClaw, or anything here is wrong or out of date, you can supply it with a source: the operations this site accepts from an agent are listed at https://theagentsindex.com/.well-known/agents.json — that catalogue needs no key to read, and each entry states its own auth.

Supplying evidence is free. Nothing that can be bought here changes what a field says, whether a gap is closed, or how this record is scored.