Skillget

ghostclaw

agent
Security Audit
Fail
Health Pass
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 24 GitHub stars
Code Fail
  • network request — Outbound network request in .claude/skills/add-discord/add/src/channels/discord.test.ts
  • network request — Outbound network request in .claude/skills/add-discord/add/src/channels/discord.ts
  • os.homedir — User home directory access in .claude/skills/add-discord/modify/src/config.ts
  • process.env — Environment variable access in .claude/skills/add-discord/modify/src/config.ts
  • process.env — Environment variable access in .claude/skills/add-slack/add/src/channels/slack.ts
Permissions Pass
  • Permissions — No dangerous permissions requested
Purpose
This tool is an AI agent designed to run directly on your computer. It acts as a personal assistant that connects to messaging platforms like Telegram and Slack, executing tasks autonomously using Claude AI.

Security Assessment
Overall risk: High. The tool is explicitly designed to operate with full, bare-metal system access, intentionally stripping away containerization like Docker. The code accesses the user's home directory and reads environment variables to configure third-party messaging channels. It also makes outbound network requests to APIs like Discord and Slack. Because the agent runs as a direct Node.js process and executes bash commands, any compromise could lead to full machine takeover or data theft. No hardcoded secrets were detected, but the setup relies on the user properly managing their own API keys.

Quality Assessment
The project is highly active, with its last code push occurring today. It uses the permissive MIT license and includes a comprehensive README with clear documentation. The tool currently has 24 GitHub stars, indicating an early-stage project with a small but growing user base.

Verdict
Use with caution — while well-maintained, its deliberate lack of sandboxing and full system access make it suitable only for dedicated or disposable machines.
SUMMARY

an AI that lives on your computer and does stuff for you. Public beta.

README.md

GhostClaw

GhostClaw

give it a machine. let it roam.

Stars License Node TypeScript Community

Website  ·  Why GhostClaw  ·  Install  ·  Skills  ·  Skill Repo

The OpenClaw alternative that actually works. A bare-metal fork of NanoClaw — containers stripped, full system access, Telegram-first. Give your spare computer its own accounts and let an AI agent run wild. 10 minutes to set up.

Why GhostClaw?

GhostClaw OpenClaw NanoClaw
Setup 10 min 30+ min 15 min
Codebase ~4K LOC ~500K LOC ~700 LOC
System access Full (bare metal) Full Sandboxed (Docker)
Messaging Telegram-first 50+ channels Telegram, WhatsApp, Slack
Containers None Required Required
Best for Solo devs, tinkerers, personal agents Teams, enterprises Security-first setups

NanoClaw's simplicity. OpenClaw's freedom. None of the mess.

GhostClaw started as a NanoClaw fork with one question: what if you ripped out everything that makes self-hosting painful? No Docker, no container orchestration, no config maze. Agents run as direct Node.js child processes with full machine access — bash, files, web, email. That's the point.

Get started

You need a Claude account (Max or API keys) and a computer to run it on. Best on a dedicated machine — a Mac Mini, an old laptop, a VPS.

npm install -g @anthropic-ai/claude-code
git clone https://github.com/b1rdmania/ghostclaw.git
cd ghostclaw && npm install
claude
# then type: /setup-ghostclaw

The setup skill walks you through everything — Telegram bot, API keys, personality, service management. About 10 minutes.

Requirements: Node.js 20+, macOS or Linux, Claude Max or API keys.

What's included

Everything below ships with every install.

Channels

  • Telegram — DM it like a person. In groups, responds when mentioned. Voice notes supported. This is the primary interface.

Core capabilities

  • Full machine access — bash, files, web search, email. No sandbox, no permission prompts.
  • Web research — Perplexity-powered search and deep research. Ask anything current, it searches and cites sources.
  • Ralph loops — autonomous multi-task engine. Hand it a checklist, it works through tasks one by one overnight.
  • Scheduled tasks — "check Hacker News every morning" or cron syntax. Natural language or precise.
  • Per-group personality — each chat gets its own CLAUDE.md defining tone, memory, and rules.
  • Model selection — switch between Sonnet, Opus, Haiku from Telegram. Just type /model.

Mission Control

Built-in dashboard at localhost:3333. Real-time activity feed, task scheduling, soul editing, research runs.

Mission Control dashboard

Architecture

You (Telegram) → GhostClaw → Claude (Agent SDK) → Response

One process. SQLite state. Claude runs as a direct child process. No containers, no Docker, no Kubernetes. Full system access — that's the point.

Optional skills

Add capabilities with a command. Each skill is a markdown file — security-scanned before install.

Command What it adds
/add-gmail-agent Gmail read/send — verification codes, urgent flags, summaries
/add-voice-transcription Voice note transcription (ElevenLabs Scribe)
/add-voice-reply Bot replies with voice notes (ElevenLabs TTS)
/add-heartbeat Periodic health checks — disk, logs, services
/add-morning-briefing Daily or weekly briefings
/add-slack Slack as an additional channel
/add-telegram-swarm Multi-bot agent teams in Telegram
/pr-babysitter Automated PR monitoring — CI fixes, review resolution
/run-ralph Autonomous overnight task loop
/update-ghostclaw Safe update: backup, pull, migrate, rebuild, restart

Browse all 46 skills → including marketing, SEO, design, and community-built skills.

Build your own — skills are markdown files. Ask Claude to write one or create it yourself.

Telegram commands

Command What it does
/reset Kill stalled agent, wipe session. Next message starts fresh.
/status Active agents, queue depth, uptime.
/model View or switch AI model (sonnet, opus, haiku).
/skills List installed skills.
/update Pull latest code, rebuild, restart — no SSH needed.
/ping Check if the bot is online.

Updating

Send /update in Telegram. No SSH required.

For a full safe update with backup and rollback tag:

/update-ghostclaw

On a version before v0.6.0? Update manually:

cd ghostclaw
git pull origin main
npm install
npm run build
# macOS: launchctl kickstart -k gui/$(id -u)/com.ghostclaw
# Linux: systemctl --user restart ghostclaw

Configuration

All config lives in .env. The setup wizard creates this.

Variable Required Description
CLAUDE_CODE_OAUTH_TOKEN Yes* Claude Max subscription token
ANTHROPIC_API_KEY Yes* Or use an API key instead
ASSISTANT_NAME Yes Bot name (trigger word in groups)
TELEGRAM_BOT_TOKEN Yes From @BotFather
GHOSTCLAW_MODEL No Default: sonnet. Also: opus, haiku
ELEVENLABS_API_KEY No For voice transcription and replies
TELEGRAM_ONLY No Set true to skip WhatsApp
GMAIL_MCP_ENABLED No Set 1 for email integration

*One of CLAUDE_CODE_OAUTH_TOKEN or ANTHROPIC_API_KEY is required.

FAQ

What does it cost?
Your Claude subscription (Max or API) and optionally ElevenLabs for voice. No platform fees.

Is this secure?
The bot has full access to its machine. That's the design — run it on dedicated hardware with fresh accounts, not your daily driver. Skills are security-scanned before install.

Can I run it on a Raspberry Pi?
Untested but should work on any Linux ARM/x64 box with Node.js 20+.

Community

Join the GhostClaw community on Telegram to share problems, suggestions, or see what others are building.

Credits

GhostClaw wouldn't exist without the work of others.

NanoClaw by qwibitai — the foundation. GhostClaw is a fork of NanoClaw. The core architecture, agent runner, skills engine (three-way merge, manifest system, replay/rebase), and the skill-based extensibility model are all their work. We stripped out containers and built on top, but the bones are theirs.

OpenClaw — the inspiration. Heartbeat monitoring, daily briefings, autonomous task loops, the idea of an agent with its own identity and accounts — these came from watching what OpenClaw pioneered.

NanoClaw contributors — Alakazam03, tydev-new, pottertech, rgarcia, AmaxGuan, happydog-intj, bindoon — whose contributions to NanoClaw's codebase carry forward into GhostClaw.

Licence

MIT

Reviews (0)

No results found