
Openclaw
github.com/openclaw/openclaw- Category
- AI Agents
- Rank
- No. 03Tools index
- Type
- APP
- Builder
- openclaw
- GitHub
- 386.5k stars
- Latest release
- pr-124528-profiles
- Added
- Apr 27, 2026
About
A self-hosted personal AI assistant that connects messaging apps to an extensible agent runtime. It runs tools and reusable skills, keeps persistent context, and can operate on infrastructure you control.
What it does
OpenClaw is a self-hosted gateway that puts an AI agent behind the chat apps you already use. One long-running service owns sessions, routing, tool permissions and channel connections, and a command-line tool, a browser control panel, a terminal interface and paired phone or desktop devices all attach to it. Roughly thirty messaging integrations, from Discord and Slack to WhatsApp and iMessage, ship mostly as installable plugins, and an OpenAI-compatible API on the same port lets existing chat frontends point straight at it.
Why it's ranked here
The integration job here is genuinely hard: around thirty messaging platforms, device pairing, sessions, scheduled jobs, memory, and MCP both served and consumed, delivered as MIT-licensed code held by a non-profit foundation rather than a shim in front of a hosted product. The OpenAI-compatible endpoints mean existing frontends work without a bespoke adapter, and declared schema numbers imply a real migration path rather than ad-hoc upgrades. The documentation reads like operational runbooks, down to port precedence tables and upgrade ordering for device fleets.
What's good
The privilege model is worked out much further than usual for a self-hosted assistant. Command approvals build a canonical run plan before the human signs off, then execute that stored plan rather than any later edited command or working directory, and bind a single concrete file operand, refusing outright when one cannot be identified. Approval scope escalates by command class, up to a distinct administrator level for host execution. One safety check, refusing to manage a user service when a same-named system daemon exists, fails closed with no force flag. The docs even carry metadata telling agents when each page is worth reading.
Tradeoffs
Surface area is the cost. The command-line tool documents more than forty top-level commands before plugins add their own, and the list of files published to the package registry is a hand-maintained deny-list of several hundred entries, a lot of manifest to keep correct. By default, tools run directly on the host for the main session unless you configure sandboxing, so the out-of-the-box posture for a channel strangers can message is host execution. The recommended install pipes a remote script into a shell, and the docs list separate workarounds for several package-manager versions, a sign of how much the install path leans on lifecycle scripts.
How to use it well
Treat it as an operator tool with a security boundary, not an appliance. Keep the gateway bound to the local machine and reach it over a private network or SSH tunnel; inside a detected container the default binding opens to all interfaces, so check that case. Supply the access token through the environment rather than a config file, and configure sandboxing before connecting any channel unknown senders can reach. Run one gateway per machine even when hosting several agents; if you must run more, give each its own port, config and state. Telegram is the easiest first channel, needing only a bot token and no plugin install.
Technical notes+
The Gateway multiplexes everything onto one port, 18789 by default: WebSocket control and RPC, plugin HTTP routes, the Control UI, and an OpenAI-compatible surface serving /v1/models, /v1/embeddings, /v1/chat/completions and /v1/responses, documented in docs/gateway/index.md; the model list returns one identifier per configured agent. Bind mode defaults to loopback but becomes auto, resolving to all interfaces, inside a detected container. Auth is required by default via token or password, configurable through OPENCLAW_GATEWAY_TOKEN or OPENCLAW_GATEWAY_PASSWORD. src/index.ts is a thin executable that installs unhandled-rejection and uncaught-exception handlers and forwards to the CLI bootstrap, deferring library exports; it marks loadSessionStore and saveSessionStore as deprecated. Nodes connect over the same socket with a signed device identity; a pairing record fixes the command surface a node may expose, pending pairing requests expire five minutes after the device's last retry, and the Gateway accepts N-1 node protocol versions. docs/nodes/index.md specifies exec approvals that canonicalize a run plan before approval, re-validate cwd, and bind exactly one local file operand. package.json declares state schema 6 and agent schema 16, and its files list is a manually maintained set of several hundred negated build-output entries. State uses node:sqlite, so a Bun-installed executable still needs a Node runtime. docs/channels/index.md is the generated channel catalog; doc pages carry read_when frontmatter for agents.
Observed
- License
- MIT, declared in the package manifest
- Governance
- Authored by the OpenClaw Foundation, a non-profit
- Packaging
- Published to npm; building from source uses a pnpm workspace, and plain npm install at the repository root is documented as unsupported
- Interfaces
- CLI, browser control UI, terminal UI, paired devices over WebSocket, and an OpenAI-compatible HTTP API
- Gateway auth
- Required by default; token or password, supplied via config or environment variables
Read from README.md, package.json, src/index.ts, docs/index.md, docs/install/index.md, docs/cli/index.md, docs/gateway/index.md, docs/tools/index.md, docs/channels/index.md, docs/nodes/index.md.
What it can do
Respond to messages as personal AI assistant
User messages via messaging platforms → AI-generated responses
Process voice commands and queries
Voice input on macOS/iOS/Android → Spoken responses and actions
Convert speech to text
Audio/voice input → Text transcription
Convert text to speech
Text responses → Audio/voice output
Render interactive live canvas
User commands and data → Visual canvas display
Connect to multiple messaging platforms
Platform credentials and configuration → Cross-platform messaging integration
Run locally on user devices
Device installation and setup → Local AI assistant instance
Intel on Openclaw
- How OpenClaw's heartbeat bug drove auto-router adoptionAug 11, 2026
- Quoting OpenClawAug 10, 2026
- "Do a full end-to-end QA test of OpenClaw with live API keys. Use 12 subagents tJul 28, 2026
- From skeptic to true believer: How OpenClaw changed my life | Claire VoJul 21, 2026
- OpenClaw install as software 3.0 patternJul 17, 2026
Tags
Tech Stack
Featured in Intel
Comments (0)
No comments yet
Indexed by a proprietary survey. Corrections welcome.