claude-telegram

A simple and modular Telegram orchestrator on top of Claude Code CLI.

One npm package that connects a Telegram bot to Claude Code via --resume sessions, with whitelist access control and live activity status.

npm install -g claude-telegram

claude-telegram vs OpenClaw

	claude-telegram	OpenClaw
Core	~1500 LOC, one dependency (grammY)	50+ integrations, large codebase
Approach	Minimal orchestrator — Claude Code does the work	Full-featured AI assistant platform
Extensibility	Module system — add anything you need	Built-in, growing feature set
Control	You own the code, easy to audit and modify	Community-driven, fast-moving
Setup	npx claude-telegram start	Multi-step setup

Both are valid choices. claude-telegram is for those who prefer a small, predictable core that they extend themselves.

Installation

1. Get a server — a VPS, a home PC, anything with internet access.

2. Install Claude Code:

curl -fsSL https://claude.ai/install.sh | bash

3. Install claude-telegram:

npm install -g claude-telegram

Or run without installing via npx claude-telegram.

Alternatively, give Claude the link and let it handle everything:

claude "install claude-telegram from github.com/bluzir/claude-telegram and set it up for my Telegram bot"

Quick Start

mkdir my-agent && cd my-agent
echo "You are a helpful assistant." > CLAUDE.md

cat > claude-telegram.yaml << 'EOF'
token: ${MY_BOT_TOKEN}
workspace: .
whitelist: [YOUR_USER_ID]
permission_mode: acceptEdits
EOF

export MY_BOT_TOKEN="123456:ABC-DEF..."
npx claude-telegram start

Don't know your Telegram user ID? Run npx claude-telegram whoami and send a message to the bot.

Config

Create claude-telegram.yaml in your project root:

token: ${TELEGRAM_BOT_TOKEN}        # env var interpolation
workspace: /path/to/workspace        # cwd for Claude CLI

# Who can use the bot (Telegram user IDs)
# Empty list = NO ONE (secure by default)
whitelist:
  - 16643982

# What Claude can do
# default | acceptEdits | bypassPermissions
permission_mode: acceptEdits

# --- Optional ---
# claude_path: /usr/local/bin/claude   # default: "claude" from PATH
# timeout: 300                          # seconds, default: 300
# model: sonnet                         # model override
# system_prompt: "Reply in Russian"     # injected into every call
# add_dirs:                             # additional dirs for Claude
#   - /path/to/shared/data
#
# --- Optional: multi-instance ---
# session_namespace: my-bot               # unique seed for session IDs (required when
#                                         # multiple bots share the same whitelist user)
#
# --- Optional: security/capabilities (advanced) ---
# disable_slash_commands: true            # disable Claude Code "skills" (slash commands)
# setting_sources: ["user", "project"]    # ignore local settings in workspace (".claude/settings.local.json")
# strict_mcp_config: true                 # disable MCP unless explicitly configured
# mcp_config:                             # MCP server configs (paths or JSON strings)
#   - ./mcp.json
#
# tools: ["Read", "Grep", "Glob"]         # restrict built-in tools ("" disables all tools)
# allowed_tools:                          # allowlist with optional patterns (e.g. Bash(git:*))
#   - "Read"
# disallowed_tools:                       # explicit denylist
#   - "Bash"

# modules:                              # optional plugin modules (loaded at startup)
#   - import: ./modules/voice.mjs       # resolved relative to `workspace`
#     options:
#       provider: openai
#   - import: claude-telegram-whoop-module
#     options:
#       client_id: ${WHOOP_CLIENT_ID}

Environment variables are interpolated with ${VAR_NAME} syntax.

Workspace

The workspace field in config is just the working directory (cwd) for Claude CLI. It can be any directory — claude-telegram doesn't care what's inside. Claude Code will use whatever CLAUDE.md, .claude/agents/, .claude/skills/, and other project files it finds there, exactly as it would in the terminal.

Simple setup

The simplest workspace is a directory with just a CLAUDE.md:

my-agent/
├── CLAUDE.md                    # Instructions for Claude
├── claude-telegram.yaml         # Bot config
└── data/                        # Runtime data (auto-created)

Using an existing project

Any project that works with Claude Code works with claude-telegram — just point workspace at it:

# Clone any project that has CLAUDE.md / .claude/ setup
git clone https://github.com/bluzir/claude-pipe

# claude-telegram.yaml
token: ${MY_BOT_TOKEN}
workspace: ./claude-pipe/examples/research-pipeline
whitelist: [YOUR_USER_ID]
permission_mode: acceptEdits

Claude will see the project's CLAUDE.md, agents, skills, commands, MCP servers — everything. You don't need to copy or restructure anything.

Multi-agent setup

Run multiple bots, each pointing to its own project:

# researcher.yaml
token: ${RESEARCHER_BOT_TOKEN}
workspace: ./research-pipeline
whitelist: [YOUR_USER_ID]
permission_mode: acceptEdits
timeout: 600
system_prompt: "You are a research agent. Be thorough and cite sources."
add_dirs:
  - ./shared

# assistant.yaml
token: ${ASSISTANT_BOT_TOKEN}
workspace: ./assistant
whitelist: [YOUR_USER_ID]
permission_mode: acceptEdits

npx claude-telegram start --config researcher.yaml
npx claude-telegram start --config assistant.yaml

Each bot gets its own workspace, sessions, and Telegram token. They share nothing unless you explicitly use add_dirs.

Modules

You can extend the bot without bloating the core by adding optional modules that register extra handlers (voice/video, API integrations, etc.).

Module import rules:

• Relative paths are resolved against workspace
• Anything else is treated as a package specifier and resolved by Node
• Modules must export a default object or a default factory function

Minimal module example ({workspace}/modules/hello.mjs):

export default function createModule() {
  return {
    name: "hello",
    commands: [{ command: "/hi", description: "Say hi" }],
    register({ bot, dispatchToClaude }) {
      bot.command("hi", async (ctx) => {
        await dispatchToClaude(ctx, "Say hi in one sentence.");
      });
    },
  };
}

Hooks (memory, security, post-processing)

Modules can also hook into the request pipeline:

• beforeClaude(ctx, message) — deny or transform the user's message before it is sent to Claude
• afterClaude(ctx, result) — observe/transform Claude result before it is sent back to Telegram

Security example (deny messages containing a secret keyword):

export default function createModule() {
  return {
    name: "security",
    async beforeClaude(ctx, message) {
      if (message.includes("OPENAI_API_KEY")) {
        return { action: "deny", reply: "Denied: looks like a secret." };
      }
      return { action: "continue" };
    },
  };
}

Memory-ish example (prepend extra context):

export default function createModule() {
  return {
    name: "memory",
    async beforeClaude(ctx, message) {
      const extraContext = "Context: you are talking to the same user as before.";
      return { action: "continue", message: `${extraContext}\n\n${message}` };
    },
    async afterClaude(ctx, result) {
      // Store result.output somewhere if you want (file/db/vector store).
      return result;
    },
  };
}

CLI

npx claude-telegram start                    # uses claude-telegram.yaml in CWD
npx claude-telegram start --config ./my.yaml # custom config path
npx claude-telegram check                    # validate config + claude CLI
npx claude-telegram whoami                   # get your Telegram user ID

Commands

Command	Description
/start	Welcome message
/cancel	Stop current request
/clear	Reset conversation (new session)
/help	Show available commands

How It Works

• Each user gets a persistent Claude session via --resume <sessionId>
• Sessions survive bot restarts (stored in ~/.claude/ by Claude CLI)
• Session mapping stored in {workspace}/data/.claude-telegram/sessions.json
• Bot responds in private chats only (ignores group/supergroup/channel)
• One message at a time per user (concurrent messages get "Still working..." reply)
• Live activity status shows what Claude is doing (reading, editing, searching, etc.)
• Messages are split at 3800 chars and formatted as Telegram MarkdownV2

Programmatic API

import { createBot, loadConfig } from "claude-telegram";

// From config file
const config = loadConfig("./claude-telegram.yaml");

// Or build config directly
const bot = createBot({
  token: process.env.BOT_TOKEN!,
  workspace: "/path/to/workspace",
  whitelist: [16643982],
  permissionMode: "acceptEdits",
  claudePath: "claude",
  timeout: 300,
});

await bot.start();

What's NOT Included

This is intentionally minimal. Not included:

• Voice/photo messages
• Multi-bot routing or gateway
• Approval system (use Claude CLI's --permission-mode)
• Budget tracking
• Web dashboard
• Queue system (one message at a time, extras are rejected)

Any of these can be added as a module without touching the core.

Security

Basic security is built into the core — whitelist, permission modes, tool restrictions, MCP lockdown, error sanitization. See SECURITY.md for details.

For advanced security (sandboxing, DLP, audit logging, and more), check out Radius.

License

MIT