What it is
OpenClaw is an open-source personal AI assistant (MIT, 375k stars, built by Peter Steinberger) that you actually chat with from the apps you already use — WhatsApp, Telegram, Slack, Discord, iMessage, Signal, Microsoft Teams and a long tail of others — while the agent itself runs locally on your laptop. The pitch is simple: instead of opening one more web tab to talk to ChatGPT, you DM your assistant where you already have conversations, and it can act on your behalf around the clock.
The product is not a wrapper around an LLM API. It is a full local-first daemon (the “Local Gateway”) that owns sessions, channels, tools, voice, scheduling and visual output, and lets you bring your own model — Claude, GPT, Gemini, or a local model via Ollama / LM Studio. Your data stays on your hardware unless you choose otherwise.
Architecture
OpenClaw is built around one idea: a single local control plane, many channels, many isolated agents.
+---------------------------------------+
| OpenClaw Local Gateway |
| (sessions, events, tools, scheduler) |
+---------------------------------------+
| | | |
v v v v
WhatsApp Telegram Slack iMessage ... (20+ channels)
| | | |
+---------+---------+----------+
|
v
+-------------+--------------+
| Multi-agent router |
| (per-channel / per-peer) |
+-------------+--------------+
|
+-----+-------+-------+------+
v v v v v
Agent Agent Agent Agent Agent
(work) (mom) (kids) (clients)(self)
Three pieces matter most:
- Local Gateway — one daemon on your machine. It is the single source of truth for sessions, events, installed tools, channel credentials and scheduled jobs. Nothing leaves the device unless a tool explicitly calls out.
- Multi-agent routing — inbound traffic is mapped to isolated agents: a different account, person, group or channel can route to a separate agent with its own workspace, memory and system prompt. Your work Slack agent does not see your iMessage history.
- Live Canvas — an agent-driven visual workspace using the A2UI interface, so the agent can render rich UI back into the chat (tables, forms, charts, plans) instead of only walls of text.
Install
Requires Node 24 (recommended) or Node 22.19+. macOS, Windows and Linux all supported.
One-line installer (recommended):
curl -fsSL https://openclaw.ai/install.sh | bash
Or via npm / pnpm directly:
npm install -g openclaw@latest
# or
pnpm add -g openclaw@latest
openclaw onboard --install-daemon
Then start the interactive onboarding (model selection, channel linking, gateway daemon):
openclaw onboard
For developers who want to hack on it, you can clone the repo and run from source — the README walks through the Git path. The daemon installs as a launch agent on macOS / a service on Linux, so it survives logout.
Configuration
Minimal config lives at ~/.openclaw/openclaw.json. The smallest possible file just picks a model:
{
"agent": {
"model": "anthropic/claude-opus-4-7"
}
}
A more realistic setup wires up two channels, gives each its own agent, and adds a couple of tools:
{
"agent": {
"model": "anthropic/claude-opus-4-7",
"fallback_model": "openai/gpt-5"
},
"channels": {
"telegram": {
"enabled": true,
"bot_token": "$TELEGRAM_BOT_TOKEN",
"route_to": "personal-agent"
},
"slack": {
"enabled": true,
"workspace": "acme-corp",
"route_to": "work-agent"
},
"imessage": {
"enabled": true,
"route_to": "personal-agent"
}
},
"agents": {
"personal-agent": {
"system_prompt_file": "~/.openclaw/prompts/personal.md",
"tools": ["gmail", "calendar", "obsidian", "browser"]
},
"work-agent": {
"system_prompt_file": "~/.openclaw/prompts/work.md",
"tools": ["github", "linear", "browser", "shell-sandboxed"]
}
},
"voice": {
"wake_word": "hey claw",
"talk_mode": true
}
}
Secrets are read from environment variables (the $VAR syntax) so the config file itself stays safe to commit to a private dotfiles repo.
Usage examples
1) Check status of the local daemon.
$ openclaw /status
gateway: running (pid 48211)
model: anthropic/claude-opus-4-7
agents: personal-agent, work-agent
channels: telegram (ok), slack (ok), imessage (ok)
sessions: 4 active
2) From Telegram, ask the personal agent to triage your inbox. You DM your own bot:
You: clear my Gmail inbox - archive newsletters,
flag anything that needs a reply today
Claw: Scanning 312 unread...
- 248 newsletters archived
- 41 receipts filed under Finance/2026
- 23 need your reply, sorted by urgency:
1. Tax accountant (deadline tomorrow)
2. Client X - contract addendum
3. ...
Want me to draft replies for the top 3?
3) Voice + Live Canvas. Say the wake word, ask for a weekly plan:
You (voice): hey claw, plan my week around the
three deadlines in my calendar
Claw: [renders a Live Canvas in chat:
Mon-Sun grid, each block linked
to a calendar event, drag-to-edit]
Supported channels
OpenClaw ships connectors for: WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, IRC, Microsoft Teams, Matrix, Feishu, LINE, Mattermost, Nextcloud Talk, Nostr, Synology Chat, Tlon, Twitch, Zalo, Zalo Personal, WeChat, QQ, and WebChat. Each channel is enabled by flipping it on in openclaw.json and pasting in the right token or QR scan during openclaw onboard.
What’s new / version
The repo currently sits at 375k stars and 53,227 commits on main, with active weekly releases. Recent additions:
- Live Canvas with the A2UI protocol — agents can return interactive UI components, not just text.
- Voice Wake + Talk Mode — wake words on macOS / iOS and continuous voice on Android (ElevenLabs + system TTS fallback).
- Multi-agent routing — map channels, accounts, or specific peers to isolated agent workspaces with separate memory.
- 50+ integrations — Gmail, GitHub, Obsidian, Twitter, Spotify, Philips Hue, calendar, files, browser automation, shell (sandboxed).
- Skills marketplace — community skills the agent can install, plus the ability for the agent to author its own.
Why it matters / where I use it
The interesting bet here is that the assistant should live where I already live — my phone’s lock screen, my work Slack, my family WhatsApp — not in a separate web tab. Once you can text the agent the way you text a person, the activation energy of asking it to do real work collapses from “open browser, log in, paste context” to “type one line.”
For me the killer features are the local-first gateway (my client conversations never leave my laptop) and per-channel agent isolation (the agent that handles family iMessage and the agent that handles client Slack share zero memory and zero tools). On the Bot-UI & Provider Hub work I do, OpenClaw is the reference for how a chat-native, channel-agnostic personal agent should be structured.
Source
GitHub: github.com/openclaw/openclaw · Site: openclaw.ai · License: MIT.