mcp-bot

Slack bot that uses Claude 3.5 Sonnet as the LLM backend and dynamically discovers tools from MCP servers. The bot runs in Socket Mode as a Slack Assistant and can manage MCP server connections at runtime.

What It Does

Listens in Slack threads via the Bolt.js Assistant API. When a user sends a message, the bot:

1. Collects thread history
2. Queries connected MCP servers for available tools
3. Sends the message + tools to Claude 3.5 Sonnet
4. Executes any tool calls against the originating MCP server
5. Returns the final response in-thread

Also supports channel summarization (fetches last 50 messages and passes them to Claude).

The bot has a "Microsoft Certified Professional" persona — it references Clippy and Teams but otherwise works normally.

Status

Area	State
Slack transport	Socket Mode (bolt.js ^4.1.1)
LLM	Claude 3.5 Sonnet via Anthropic SDK
MCP client	@modelcontextprotocol/sdk ^1.0.3
Language	TypeScript
License	MIT
CI	Biome linting via GitHub Actions

Architecture

src/
  app.ts          — Bolt.js app, Assistant handler, tool dispatch loop
  mcp-manager.ts  — Manages MCP server lifecycle (add/remove/list)
MCP/
  dspy-docs-server/ — Bundled MCP server for DSPy documentation

The bot maintains a toolToServerMap that tracks which MCP server provides each tool. When Claude returns a tool_use block, the bot routes the call to the correct server.

MCP Server Management

Three built-in tools let you manage servers from within Slack:

Command	What it does
add_mcp_server	Registers a new MCP server by name, command, and args
remove_mcp_server	Disconnects and removes a server
list_mcp_servers	Returns all registered servers

Servers communicate over stdio. One server (dspy-docs) is added by default on startup.

Setup

npm install
npm run build
npm start

Environment Variables

Variable	Required	Description
SLACK_BOT_TOKEN	Yes	xoxb- bot token
SLACK_APP_TOKEN	Yes	xapp- app-level token for Socket Mode
ANTHROPIC_API_KEY	Yes	Anthropic API key

Slack App Configuration

1. Create app at https://api.slack.com/apps
2. Enable Socket Mode
3. Add bot scopes: chat:write, channels:history, groups:history, assistant:write
4. Install to workspace
5. Copy tokens to .env

Limitations

• Tool call loop is single-pass: if Claude returns multiple tool calls, each gets a separate follow-up LLM call rather than batching results
• Thread context is rebuilt from Slack API on every message (no local cache)
• The hardcoded dspy-docs server path points to a local filesystem path (/Users/speed/mcp-bot/MCP/...)
• No retry logic for MCP server connections
• No rate limiting on Anthropic API calls
• Channel summarization is capped at 50 messages

Dependencies

Package	Version	Purpose
@slack/bolt	^4.1.1	Slack app framework
@anthropic-ai/sdk	^0.32.1	Claude API client
@modelcontextprotocol/sdk	^1.0.3	MCP client transport
openai	^4.74.0	Listed but unused in current code
typescript	^5.3.3	Build toolchain

License

MIT