Clawdbot Open‑Source: A Siri‑Like Personal AI Assistant You Can Command From Your Phone

Definition

Clawdbot defines itself as “The AI that actually does things.” It runs locally on macOS, Windows (WSL2) or Linux, keeping all data on the device.

Architecture

Three‑layer design: Gateway (control plane, WebSocket server), Channels (chat integrations), Agent (local AI worker).

WhatsApp / Telegram / Slack / Discord / Signal / iMessage / Teams / WebChat
               │
               ▼
        ┌───────────────────┐
        │     Gateway       │
        │ (control plane)   │
        │ ws://127.0.0.1   │
        └─────────┬─────────┘
                  │
   ┌─────────────────────┼─────────────────────┐
   │                     │                     │
 Pi Agent (RPC)      CLI command line      macOS / iOS / Android

Core capabilities

1. Multi‑channel access (14+ platforms)

WhatsApp – QR‑code pairing via Baileys

Telegram – Bot token (grammY)

Discord – Bot token (discord.js)

Slack – Workspace app (Bolt)

Signal – signal‑cli

iMessage – AppleScript bridge (macOS only)

Microsoft Teams – enterprise extension

Matrix – Matrix protocol

Nextcloud Talk – self‑hosted

Nostr – encrypted DM (NIP‑04)

Zalo – Bot API / QR login

WebChat – built‑in UI

2. Multi‑model support

Anthropic – Claude Pro/Max + Opus 4.5 (recommended)

OpenAI – GPT‑4, GPT‑5, o1

Google – Gemini 2.5 Pro/Flash

xAI – Grok 3 & 4

DeepSeek – V3 & R1

Mistral – Large & Codestral

GLM – ChatGLM series

OpenRouter – unified API gateway

Hugging Face – open‑source models

Local models – Ollama, LM Studio

3. Browser control (real RPA)

Clawdbot creates an isolated Chrome/Chromium profile named clawd and drives it via Chrome DevTools Protocol (CDP) + Playwright. Supported actions include deterministic tab management, clicking, typing, form filling, taking screenshots, exporting PDFs, and managing multiple profiles.

# View browser status
clawdbot browser --browser-profile clawd status
# Start browser
clawdbot browser --browser-profile clawd start
# Open a page
clawdbot browser --browser-profile clawd open https://example.com
# Take a snapshot for AI consumption
clawdbot browser --browser-profile clawd snapshot
# Click an element (ref 12)
clawdbot browser click 12
# Type into a field
clawdbot browser type 23 "hello world"
# Wait for page state
clawdbot browser wait --url "**/dashboard"
clawdbot browser wait --load networkidle
clawdbot browser wait "#main-content"

4. System command execution (Exec tool)

The Exec tool runs shell commands in a sandboxed Docker container (default) or on the host. Security modes are deny (default), allowlist, and full. Example configuration:

{
  "tools": {
    "exec": {
      "host": "sandbox",
      "security": "allowlist",
      "ask": "on-miss",
      "timeout": 1800,
      "pathPrepend": ["~/bin", "/opt/oss/bin"]
    }
  }
}

Long‑running commands can be started and polled; key presses can be sent to a running process.

5. Skill system

Skills are defined by a SKILL.md manifest and can reside in three locations: bundled with the package, ~/.clawdbot/skills (managed/local), or the workspace skills folder. The public market clawdhub.com hosts community‑contributed skills that can be installed with a single CLI command.

---
name: nano-banana-pro
description: Generate or edit images via Gemini 3 Pro Image
metadata: {"clawdbot":{"requires":{"bins":["uv"],"env":["GEMINI_API_KEY"]},"primaryEnv":"GEMINI_API_KEY"}}
---

Usage instructions...

6. Productivity integrations

Notes: Apple Notes, Obsidian, Bear, Notion

Task managers: Apple Reminders, Things 3, Todoist, Trello

Code: GitHub issues/PRs

Music: Spotify, Sonos

Smart home: Philips Hue, Home Assistant, 8Sleep

Automation: cron, webhooks, Gmail Pub/Sub

Media utilities: AI image generation, GIF search, screenshots, camera

Password management: 1Password

Weather queries

7. Voice capabilities

Voice wake‑up (“Hey Siri” style)

Continuous talk mode

Optional speech synthesis via ElevenLabs

Installation & onboarding

One‑click installers:

macOS / Linux: curl -fsSL https://clawd.bot/install.sh | bash Windows PowerShell: iwr -useb https://clawd.bot/install.ps1 | iex npm / pnpm: npm install -g clawdbot@latest or pnpm add -g clawdbot@latest The onboarding wizard ( clawdbot onboard --install-daemon) guides the user through:

Selecting a local or remote Gateway

Configuring an AI model (OAuth for Claude/OpenAI or API key)

Pairing chat channels (e.g., WhatsApp QR code, Telegram bot token)

Installing a system service (launchd or systemd)

Security model

DM pairing: unknown contacts receive a pairing code that must be approved ( clawdbot pairing approve …). An open policy can be enabled with dmPolicy="open" and an allowlist entry of *.

Sandboxing: the main session runs with full host privileges; non‑main sessions (e.g., group chats) can be forced into a Docker sandbox via agents.defaults.sandbox.mode: "non-main". Default sandbox whitelist: bash, process, read, write, edit, sessions_*. Default blacklist: browser, canvas, nodes, cron, discord, gateway. Dangerous tools can require explicit approval with ask: "always".

Remote access

Clawdbot can run on a Linux server and be accessed via Tailscale or SSH tunnels. Two modes are supported: serve – internal HTTPS using Tailscale identity. funnel – public HTTPS with password authentication.

Configuration files

Primary configuration file: ~/.clawdbot/clawdbot.json. Minimal example:

{
  "agent": {
    "model": "anthropic/claude-opus-4-5"
  }
}

Full configuration includes browser settings, tool security, channel tokens, and skill entries (see source for details).

Example community use cases

🛒 Tesco automatic shopping – weekly menu → cart → delivery slot → order, driven entirely by browser automation.

🍷 Wine cellar management – a CSV of 962 bottles turned into a local skill in minutes.

📱 iOS app development – a map‑enabled iOS app built and deployed via Telegram without opening a computer.

💻 “Sofa development” – personal site migrated from Notion to Astro while watching Netflix.

🖨️ Bambu 3D printer control – status, job queue, camera view, material management, calibration via chat.

💍 Oura health assistant – combines ring data, calendar, and fitness plan for personal health advice.

🧹 Roborock robot control – natural‑language commands to start cleaning.

🏠 Home Assistant integration – voice‑driven smart‑home actions.

🌤️ Morning briefing – generates a visual summary with weather, tasks, date, and quotes.

🔍 PR code review → Telegram – AI reviews diffs and reports merge readiness.

🗣️ Chinese learning engine – pronunciation feedback and lesson flow via voice.

👥 “Dream Team” – 14+ agents coordinated by Opus 4.5 (orchestrator) and Codex (executor).

Chat commands (example)

/status

– show session status, model usage, token consumption. /new or /reset – reset the conversation. /compact – compress context and generate a summary. /think <level> – set thinking level (off/minimal/low/medium/high/xhigh). /verbose on/off – toggle detailed mode. /usage off/tokens/full – control usage display per reply. /activation mention/always – set group activation mode. /restart – restart the gateway (admin only).

Conclusion

Clawdbot provides a local‑first personal AI assistant with deep integration across dozens of chat platforms, real‑time browser RPA, sandboxed system command execution, and an extensible skill marketplace. Its architecture and security design make it suitable for developers who need a highly customizable AI agent that can perform concrete actions on the host system.