claude-tg

Full Claude Code — in Telegram. Not a chatbot wrapper. The real agentic CLI with tools, MCP servers, and file access, streamed to your phone with live token updates.

pip install claude-tg

Why this exists

Every Telegram bridge for Claude spawns a new process per message and scrapes stdout. That means cold starts, no context between turns, and no way to interact while Claude is thinking.

claude-tg keeps a single persistent Claude Code process alive using the official --input-format stream-json protocol. The same streaming interface the terminal uses internally.

What this gives you:

• Zero startup delay — process is already running, first token appears instantly
• Mid-turn messages — send something while Claude is working, it arrives in-context (not queued)
• Every tool works — Bash, file I/O, MCP servers, web search — exactly like terminal
• Real streaming — token-by-token updates to Telegram, not periodic polling

~1,500 lines of Python. No framework, no database, no Docker.

Quick start

export TELEGRAM_BOT_TOKEN="..."    # from @BotFather
export TELEGRAM_CHAT_ID="..."      # your Telegram user ID
claude-tg

On first run, the built-in MCP server for file sending is auto-registered.

How it works

Telegram ←→ bot.py (python-telegram-bot)
                ↕
            runner.py (persistent subprocess)
                ↕ stdin/stdout (NDJSON stream-json)
            Claude Code CLI
                ↕
            MCP servers, tools, files

runner.py manages a single long-lived Claude process. Messages go to stdin as NDJSON. Events stream from stdout line by line. When a turn completes, the process stays alive for the next one.

A background reader continuously drains stdout into an asyncio queue, preventing pipe buffer deadlocks during mid-turn injection.

stream.py handles the Telegram side — buffering tokens, splitting long outputs into message chains at ~3800 chars, rate-limiting edits to stay within API limits.

bot.py ties it together: debouncing rapid messages, injecting mid-turn, handling media and voice.

Features

Streaming & persistence

• Token-level streaming with automatic message chaining (long responses split across messages)
• Persistent process across all conversation turns
• Session resume on reconnect
• Auto-reset after configurable inactivity timeout

Mid-turn interaction

• Send messages while Claude is working — they're injected via stdin into the running conversation
• Messages are debounced (0.5s) so rapid typing merges into one
• Cancel button on every response

Media

• Photos & documents → saved locally, passed to Claude as file references
• Voice messages → transcribed via Groq Whisper (free), then sent as text
• Files from Claude → sent back to Telegram via built-in MCP server

External automation (triggers)

export CLAUDE_TG_TRIGGER_PORT=9357

# Prompt goes through Claude
curl -d "Summarize today's git log" localhost:9357

# Bypass Claude, send directly to Telegram
curl -d "DIRECT:Deploy complete ✅" localhost:9357

Prompts queue if Claude is busy. Localhost only — no auth needed, no exposure.

Use this for cron jobs, monitoring scripts, heartbeats, CI/CD notifications.

Conversation log

All messages (user, assistant, triggers, direct) are persisted to data/conversation_log.jsonl. The built-in MCP tool get_conversation_context lets Claude read recent chat history for context continuity across sessions.

Commands

Running as a service (VPS)

uv tool install claude-tg

# /etc/systemd/system/claude-tg.service
[Unit]
Description=Claude TG
After=network-online.target

[Service]
Type=simple
EnvironmentFile=/path/to/.env
Environment=PATH=/root/.local/bin:/usr/local/bin:/usr/bin
ExecStart=/root/.local/bin/claude-tg --work-dir /root
Restart=always
RestartSec=5

[Install]
WantedBy=multi-user.target

systemctl enable --now claude-tg

Running as root? Claude Code blocks --dangerously-skip-permissions for root. claude-tg detects this and switches to --allowedTools with auto-discovered MCP servers from ~/.claude.json and .mcp.json.

Configuration

Requirements

• Python 3.11+
• Claude Code CLI installed and authenticated
• Telegram bot from @BotFather

License

MIT