claudeclaw-starter

A production-ready starter for ClaudeClaw — Claude Code as your personal AI assistant, running 24/7 via Telegram.

---

What is this?

ClaudeClaw turns Claude Code into a persistent, identity-aware assistant that lives on your server and talks to you over Telegram (or Discord). It's not a chatbot wrapper — it's Claude Code itself, with full access to your tools, files, and shell, controlled from your phone.

This starter gives you a pre-configured setup with:

• Telegram bot — chat with your assistant anywhere, send voice messages, images
• OpenRouter support — use any model (Claude, GPT, Gemini, Llama...) as primary or fallback
• Image generation — ask for an image, get one inline via Gemini or any OpenRouter image model
• Agentic mode — auto-routes messages to the right model based on intent (planning vs implementation)
• Heartbeat daemon — always-on background process, survives session resets
• Cron jobs — schedule recurring prompts (daily reports, server checks, reminders)
• Web UI — optional browser dashboard at localhost:4632
• Plugin ecosystem — one-command install for code review, browser automation, loops, hooks, and more

---

Quick start

git clone https://github.com/doomL/claudeclaw-starter my-claudeclaw
cd my-claudeclaw
chmod +x setup.sh && ./setup.sh

Copy the settings template and fill in your credentials:

cp config/claudeclaw-settings.template.json .claude/claudeclaw/settings.json

Edit .claude/claudeclaw/settings.json — at minimum set:

• telegram.token — your bot token from @BotFather
• telegram.allowedUserIds — your Telegram user ID (get it from @userinfobot)
• api — your Anthropic API key (or leave empty to use Claude Code's default auth)
• fallback.api — optional OpenRouter key for fallback/image models

Then edit CLAUDE.md — give your assistant a name, personality, and fill in your context.

claude .

---

Telegram features

Once your bot is running, you can:

What	How
Send a message	Just text it
Send a voice note	Transcribed automatically via Whisper
Send an image	Attached as context to your message
Ask for an image	"generate me a..." → image comes back inline
Schedule a task	"remind me to check the server every day at 8am"
Run a one-shot prompt	/start "summarize my git log"

---

Agentic model routing

ClaudeClaw can automatically pick the right model based on what you're asking:

"agentic": {
  "enabled": true,
  "modes": [
    { "name": "planning", "model": "opus", "keywords": ["plan", "design", "research"] },
    { "name": "implementation", "model": "sonnet", "keywords": ["code", "fix", "deploy"] }
  ]
}

Use OpenRouter models as fallback: "model": "openrouter/google/gemini-2.5-pro".

---

Plugins

After setup, install plugins inside Claude Code with /install-skill:

Plugin	What it does
ralph-loop@claude-plugins-official	Autonomous loop mode — runs prompts on repeat
hookify@claude-plugins-official	Prevent unwanted behaviors with pre-tool hooks
commit-commands@claude-plugins-official	/commit, /push, /pr shortcuts
code-review@claude-plugins-official	AI code review on your diffs
pr-review-toolkit@claude-plugins-official	Full PR review suite
dev-browser@dev-browser-marketplace	Browser automation with persistent state
plugin-dev@claude-plugins-official	Tools for building your own plugins

---

File structure

.
├── CLAUDE.md                              # Assistant identity + your personal context
├── config/
│   ├── settings.json                      # Claude Code: registers claudeclaw marketplace
│   └── claudeclaw-settings.template.json  # ClaudeClaw: credentials + behavior template
├── setup.sh                               # One-shot setup script
└── README.md

The ClaudeClaw runtime config lives at .claude/claudeclaw/settings.json inside your project — not the Claude Code settings.json. Two different files, both important.

---

Real things people have done with this

This isn't a toy. Here's what's been built by talking to the assistant from a phone:

Mobile app CI/CD — from Telegram

"build me a new APK"

The assistant SSHes into the server, runs expo eas build, monitors the build log, and sends back the .apk link when done. No laptop needed.

Server security monitoring — automated

A cron job runs every night and does exactly this:

check docker container logs for errors, failed auth attempts, and
anything that looks like an intrusion — summarize and alert me

If something looks off, it messages you on Telegram unprompted.

WordPress breach forensics

"something's wrong with the site"

The assistant dug through PHP files, found injected malware, identified a rogue admin account added via the database, cleaned it up, and hardened the config — all while narrating what it found.

Self-improvement loop

"install the hookify plugin and configure it to prevent that thing you keep doing"

The assistant installs its own plugins, edits its own CLAUDE.md, writes new skills, and evolves its behavior based on your feedback. The setup you're running right now is the result of a session like this.

Tournament apps — built and deployed from Telegram

"make me a darts tournament bracket for tonight, with a paid entry system"

Full HTML/JS single-page app, built and ready to open in a browser. No IDE open.

Content pipelines and LLM eval infrastructure

Configured n8n workflows, Qdrant vector caches, LLM evaluation prompts, Spring Boot microservices — all while explaining what each piece does and why.

---

The pattern is the same every time: you describe what you need, the assistant does the work — reading files, running commands, writing code, monitoring processes — and reports back. The only interface is Telegram.

---

Credits

Built on ClaudeClaw by moazbuilds.