Telegram Bridge MCP
Bridges AI assistants to a Telegram bot to enable two-way messaging, interactive confirmations, and live status updates. It supports automatic voice transcription via local Whisper models and provides secure, single-user communication for MCP-compatible hosts.
README
Telegram Bridge MCP
<img align="right" src="interaction.jpg" width="320" alt="AI agents coordinating through Telegram Bridge MCP" />
No Claw? No Problem.
Anthropic restricted Claude Code's native instance API — but this bridge doesn't care. It's a standard Model Context Protocol server. Any IDE, any model, any agent framework that speaks MCP connects out of the box — no proprietary lock-in, no webhooks, no public URL required.
What Is This?
Telegram Bridge MCP connects AI assistants to Telegram bidirectionally. It lets any MCP-compatible client send messages, ask questions, receive voice replies, and run multiple concurrent agent sessions — all through a single bot you control.
Works with: VS Code (GitHub Copilot Chat), Claude Code, Cursor, Windsurf, Copilot CLI, and any MCP-compatible host.
Highlights
| Feature | Description |
|---|---|
| Two-way messaging | Text, Markdown, files, voice notes |
| Interactive controls | Inline buttons, confirmations, questions |
| Super tools | Self-pinning checklists and emoji progress bars that update in-place |
| Voice | Auto-transcription (bundled Whisper ONNX, no ffmpeg) + TTS (local Kokoro, OpenAI, or bundled ONNX) |
| Multi-session | Multiple agents share one bot with isolated queues, token auth, and color identity |
| Animations | Cycling status frames while your agent works |
| Reminders | Scheduled synthetic events delivered via dequeue |
| Slash commands | Dynamic bot menu; commands arrive as structured events |
| No webhooks | Pure long-polling — no public URL, no reverse proxy |
Quick Start
Tip: If your AI has web access, paste this to get started (requires web access):
Set me up: https://github.com/electricessence/Telegram-Bridge-MCP
<details> <summary><strong>Manual setup (step by step)</strong></summary>
1. Clone and build
git clone https://github.com/electricessence/Telegram-Bridge-MCP.git
cd Telegram-Bridge-MCP
pnpm install && pnpm build
2. Create a bot
Message @BotFather on Telegram:
/newbot
Copy the token it gives you.
3. Pair interactively
pnpm pair
The wizard prompts for your bot token and Telegram user ID, writes a .env file, and verifies connectivity.
4. Configure your MCP host
See docs/setup.md for per-client config snippets (VS Code, Claude Code, Cursor, Docker).
</details>
Transports
| Transport | Entry Point | Best For |
|---|---|---|
| Streamable HTTP | pnpm start -- --http |
Multiple clients sharing one server (recommended) |
| stdio | node dist/index.js |
Single client, no persistent server |
| Launcher bridge | node dist/launcher.js |
Auto-starts HTTP if needed, bridges stdio ↔ HTTP |
<details> <summary><strong>Streamable HTTP MCP config example</strong></summary>
Claude Code / Cursor / other MCP hosts
{
"mcpServers": {
"telegram": {
"type": "streamable-http",
"url": "http://127.0.0.1:3099/mcp"
}
}
}
VS Code (.vscode/mcp.json)
{
"servers": {
"telegram": {
"type": "streamable-http",
"url": "http://127.0.0.1:3099/mcp"
}
}
}
</details>
<details> <summary><strong>stdio MCP config example</strong></summary>
Claude Code / Cursor / other MCP hosts
{
"mcpServers": {
"telegram": {
"command": "node",
"args": ["/path/to/Telegram-Bridge-MCP/dist/index.js"],
"env": {
"BOT_TOKEN": "your-token",
"ALLOWED_USER_ID": "your-user-id"
}
}
}
}
VS Code (.vscode/mcp.json)
{
"servers": {
"telegram": {
"type": "stdio",
"command": "node",
"args": ["/path/to/Telegram-Bridge-MCP/dist/index.js"],
"env": {
"BOT_TOKEN": "your-token",
"ALLOWED_USER_ID": "your-user-id"
}
}
}
}
</details>
Tools — The v6 API
Version 6 consolidates the entire API into 4 tools with type-based routing. Call help(topic?) at any time for interactive documentation discovery.
send — Outbound Messaging
All outbound operations flow through a single send call. The type parameter determines behavior.
| Type | Description |
|---|---|
text |
Formatted Markdown text; pass audio: "..." to speak via TTS |
file |
Photo, document, video, audio, or voice note |
notification |
Status notification with severity: info · success · warning · error |
choice |
Message with inline buttons (non-blocking) |
question |
Blocking prompt — route with ask, confirm, or choose |
dm |
DM to another session (target_sid or target alias); "direct" accepted as alias |
append |
Append text to an existing message |
animation |
Start a cycling status animation |
checklist |
Create a self-pinning live checklist; requires title, accepts steps array of {label, status} objects |
progress |
Create an emoji progress bar (width configurable) |
Update in-place with
action(type: "checklist/update", message_id: ...)andaction(type: "progress/update", message_id: ...)respectively. Seedocs/super-tools.md.
// Examples — token required for all session-scoped calls; session/start and most help topics work without a token
// token: 1234567 (required for all session-scoped calls)
send({ token: 1234567, type: "text", text: "Hello from your AI agent!" })
send({ token: 1234567, type: "notification", severity: "success", text: "Build passed." })
send({ token: 1234567, type: "question", ask: "Proceed with deployment?" })
send({ token: 1234567, type: "checklist", title: "Pipeline", steps: [{ label: "Design", status: "pending" }, { label: "Implement", status: "pending" }, { label: "Review", status: "pending" }, { label: "Deploy", status: "pending" }] })
send({ token: 1234567, type: "progress", title: "Processing files", percent: 40, subtext: "4 of 10 complete", width: 10 })
dequeue — Receive Inbound Events
Long-poll for the next inbound event: messages, button presses, voice notes, slash commands, reminders.
Note:
tokenis the integer returned byaction(type: "session/start").
dequeue({ token: 1234567 }) // default timeout — idles until event
dequeue({ token: 1234567, timeout: 0 }) // non-blocking drain (coordination gates)
action — Universal Dispatcher
RESTful path routing via type. Supports progressive discovery:
- Omit
type→ list all categories - Pass a category → list sub-paths
- Pass a full path → execute
<details> <summary><strong>Full action reference</strong></summary>
Session
session/start · session/close · session/list · session/rename · session/idle
Profile
profile/voice · profile/topic · profile/save · profile/load · profile/import · profile/dequeue-default
Reminder
reminder/set · reminder/cancel · reminder/list
Animation
animation/default · animation/cancel
Message
message/edit · message/delete · message/pin · message/route · message/history · message/get
Chat
chat/info
Super Tools
checklist/update · progress/update
Confirm Presets
confirm/ok · confirm/ok-cancel · confirm/yn
Standalone
react · acknowledge · show-typing · commands/set · logging/toggle · transcribe · download
Governor-only
approve · shutdown · shutdown/warn · log/get · log/list · log/roll · log/delete · log/debug
</details>
help — Documentation Discovery
help() // list all topics
help({ topic: "send" }) // targeted reference for a specific tool or type
Multi-Session
Multiple agents can share one bot simultaneously without cross-talk.
session/start → token (integer) → pass on every session-scoped call
Token format: token = sid * 1_000_000 + pin — a single integer, returned by action(type: "session/start").
| Capability | Description |
|---|---|
| Isolated queues | Per-session routing; no messages bleed between agents |
| Color identity | Outbound messages prefixed with color + name (e.g., 🟩 Worker 1) |
| Governor model | First session is primary; additional sessions require operator approval via color-picker keyboard |
| DMs | Inter-session messaging via send(type: "dm", target_sid: N, ...) (alias: "direct"; target_sid alias: target) |
| Health monitoring | Unresponsive sessions trigger operator prompts to reroute or promote |
| Graceful teardown | Orphaned events rerouted; callback hooks replaced on close |
See docs/multi-session-protocol.md for the full routing protocol.
Voice
Transcription (Inbound)
Voice messages are auto-transcribed before delivery. No external API, no ffmpeg required — the Whisper ONNX model is bundled.
WHISPER_MODEL=onnx-community/whisper-base # default
WHISPER_CACHE_DIR=/path/to/cache # optional
Text-to-Speech (Outbound)
Triggered by send(type: "text", audio: "..."). Provider is selected automatically:
| Environment Variable | Provider |
|---|---|
TTS_HOST |
Any OpenAI-compatible /v1/audio/speech endpoint |
OPENAI_API_KEY |
api.openai.com |
| Neither set | Bundled ONNX model (zero config) |
Kokoro (recommended local TTS)
High-quality local TTS with 25+ voices. No API key, no cost.
docker run -d --name kokoro -p 8880:8880 ghcr.io/hexgrad/kokoro-onnx-server:latest
TTS_HOST=http://localhost:8880
TTS_FORMAT=ogg
TTS_VOICE=af_heart
Send /voice in Telegram to browse and sample voices live.
Per-session voice override: action(type: "profile/voice") or /voice in Telegram.
MCP Resources
Five resources are available to any connected client — no tool call required:
| URI | Contents |
|---|---|
telegram-bridge-mcp://agent-guide |
Behavioral guide for AI agents |
telegram-bridge-mcp://communication-guide |
Communication patterns and loop rules |
telegram-bridge-mcp://quick-reference |
Hard rules + compact tool table |
telegram-bridge-mcp://setup-guide |
Setup walkthrough |
telegram-bridge-mcp://formatting-guide |
Markdown / MarkdownV2 / HTML reference |
Docker
ghcr.io/electricessence/telegram-bridge-mcp:latest
Before running Docker: Create your
.envfile first by runningpnpm pairon a machine with Node.js, or copy.env.exampleand fill it in manually.
Streamable HTTP (recommended) — run as a long-lived service:
docker run -d --name telegram-mcp \
--env-file /absolute/path/to/.env \
-e MCP_PORT=3099 \
-p 3099:3099 \
-v telegram-mcp-cache:/home/node/.cache \
ghcr.io/electricessence/telegram-bridge-mcp:latest
Connect MCP hosts to http://127.0.0.1:3099/mcp.
<details> <summary><strong>stdio mode</strong> (per-host process, no persistent server)</summary>
{
"command": "docker",
"args": [
"run", "--rm", "-i",
"--env-file", "/absolute/path/to/.env",
"-v", "telegram-mcp-cache:/home/node/.cache",
"ghcr.io/electricessence/telegram-bridge-mcp:latest"
]
}
</details>
The cache volume persists Whisper and TTS model weights across container restarts.
Development
pnpm build # Compile TypeScript
pnpm dev # Watch mode
pnpm test # Run tests
pnpm coverage # Coverage report
pnpm pair # Re-run pairing wizard
Agent Setup
To keep agents reliably in the Telegram dequeue loop, install the loop-guard hook for your host. The hook prevents agents from dropping out of the loop on idle or forced stop.
See docs/agent-setup.md for installation instructions for VS Code (GitHub Copilot Chat) and Claude Code.
Documentation
| Doc | Contents |
|---|---|
docs/setup.md |
Full setup walkthrough with per-client config |
docs/multi-session-protocol.md |
Multi-session routing and governor model |
docs/super-tools.md |
Checklist and progress bar reference |
docs/agent-setup.md |
Loop-guard hooks for VS Code and Claude Code |
docs/migration-v5-to-v6.md |
v5 → v6 tool name mapping |
docs/git-index-safety.md |
Git index safety notes for multi-agent environments |
License
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。