Band MCP Server

A Model Context Protocol (MCP) server that provides seamless integration with the Band AI platform. Enable AI agents to interact with Band's agent management, chat rooms, and messaging systems.

✨ Features

• Dual-scope tool surface: serve agent tools (--scope agent), human tools (--scope human), or both
• Opt-in contact directory (--tools contacts) and memory (--tools memory) tool groups
• Room pinning with --room-id — hides the room field from the advertised schema and injects it at call time
• STDIO transport for IDE integration; SSE transport for Docker and remote deployments
• Tool definitions sourced from band-sdk so the MCP stays in lockstep with the platform SDK

Migrating from pre-v1.2.0

Every tool name changed. Tools are now prefixed with band_, and the agent surface was reshaped when the handwritten handlers were deleted in favor of the SDK-driven registrar. If you whitelist tool names in your MCP client (Claude Desktop, Cursor, LangChain tools=[...]), expect breakage until you update them.

Notable behavior changes:

• Contact tools are no longer registered by default. Pass --tools contacts to restore them.

• get_agent_me, list_agent_chats, and message-lifecycle tools (mark_agent_message_*) have been removed. AgentTools is room-scoped via the SDK; agent identity travels with the credential.

• A handful of agent tools were renamed beyond the prefix (create_agent_chat → band_create_chatroom, list_agent_peers → band_lookup_peers, etc.).

• All THENVOI_* environment variables have been dropped with no fallback — set the BAND_* equivalent before upgrading, or the server starts with empty credentials (ConfigError at best, 401s at worst):

  Old (THENVOI_*)	New (BAND_*)
  THENVOI_API_KEY	BAND_API_KEY
  THENVOI_BASE_URL	BAND_BASE_URL
  THENVOI_USER_KEY	BAND_USER_KEY
  THENVOI_AGENT_KEY	BAND_AGENT_KEY
  THENVOI_MCP_SCOPE	BAND_MCP_SCOPE
  THENVOI_MCP_TOOLS	BAND_MCP_TOOLS
  THENVOI_MCP_ROOM_ID	BAND_MCP_ROOM_ID

🚀 Quick Start

Prerequisites

• Python 3.11 or higher
• Band API key from app.band.ai/settings/api-keys

Install from PyPI

pip install band-mcp
# or, if you use uv
uv tool install band-mcp

This installs the band-mcp CLI on your PATH. No repo clone, no uv directory flags, no absolute paths required.

Getting Your API Key

1. Log in to Band
2. Navigate to Settings → API Keys
3. Click Create New API Key
4. Copy the key immediately (won't be shown again)

📦 Install in Your IDE

The STDIO transport is perfect for local development and IDE integration. The server starts automatically when your AI assistant needs it.

IDE Integration

Configure your AI assistant to use the Band MCP Server with the following JSON structure:

{
  "mcpServers": {
    "band": {
      "command": "band-mcp",
      "args": [
        "--scope",
        "agent,human",
        "--tools",
        "contacts"
      ],
      "env": {
        "BAND_AGENT_KEY": "band_a_your_agent_key",
        "BAND_USER_KEY": "band_u_your_user_key",
        "BAND_BASE_URL": "https://app.band.ai"
      }
    }
  }
}

Note: This assumes band-mcp is installed via pip or uv tool install so the band-mcp command is on your PATH. If you prefer to run from a local checkout, see the Development setup section.

Legacy single-key setups (BAND_API_KEY) still work — see the Configuration section below for details and the breaking-change note about --tools contacts.

{
  "claude.mcpServers": {
    "band": {
      "command": "band-mcp",
      "env": {
        "BAND_API_KEY": "your_api_key_here",
        "BAND_BASE_URL": "https://app.band.ai"
      }
    }
  }
}

Manual Testing (STDIO)

For testing or standalone usage without an IDE:

# After installing band-mcp from PyPI
BAND_API_KEY=your-key band-mcp

# Or, from a local checkout
uv run band-mcp

Expected output:

2025-11-19 17:09:51,621 - band-mcp - INFO - Starting band-mcp-server v1.0.0
2025-11-19 17:09:51,621 - band-mcp - INFO - Base URL: https://app.band.ai
2025-11-19 17:09:51,621 - band-mcp - INFO - Server ready - listening for MCP protocol messages on STDIO

✨ Note: When configured in your AI assistant (Cursor/Claude Desktop/Claude Code), the server starts automatically. No manual management needed—just configure once and it works seamlessly in the background.

SSE Transport Mode (Remote/Docker Deployments)

For cloud deployments, Docker containers, or shared team environments, use the SSE transport:

# Start SSE server on default port 8000
band-mcp --transport sse

# Custom host and port
band-mcp --transport sse --host 0.0.0.0 --port 3000

Expected output:

2025-12-18 17:15:55 - band-mcp - INFO - Starting band-mcp-server v1.0.0
2025-12-18 17:15:55 - band-mcp - INFO - Base URL: https://app.band.ai
2025-12-18 17:15:55 - band-mcp - INFO - Transport: SSE (HTTP server mode)
2025-12-18 17:15:55 - band-mcp - INFO - Server ready - listening on http://127.0.0.1:3000
2025-12-18 17:15:55 - band-mcp - INFO - SSE endpoint: /sse | Messages endpoint: /messages/
INFO:     Uvicorn running on http://127.0.0.1:3000 (Press CTRL+C to quit)

Testing SSE Mode with curl

SSE requires maintaining a persistent connection. Use three terminals:

Terminal 1 - Start the server:

band-mcp --transport sse --port 3000

Terminal 2 - Connect to SSE stream (keep running):

curl -N http://127.0.0.1:3000/sse

You'll receive a session ID:

event: endpoint
data: /messages/?session_id=abc123def456...

Terminal 3 - Send requests (use the session ID from Terminal 2):

# 1. Initialize the connection (required first)
curl -X POST "http://127.0.0.1:3000/messages/?session_id=YOUR_SESSION_ID" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}'

# 2. List available tools
curl -X POST "http://127.0.0.1:3000/messages/?session_id=YOUR_SESSION_ID" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}'

# 3. Call a tool (e.g., health_check)
curl -X POST "http://127.0.0.1:3000/messages/?session_id=YOUR_SESSION_ID" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"health_check","arguments":{}}}'

Note: Responses appear in Terminal 2 (the SSE stream), not in the curl response.

Environment Variables for SSE

You can also configure via environment variables:

export TRANSPORT=sse
export HOST=0.0.0.0
export PORT=3000
band-mcp

Testing with MCP Inspector

npx @modelcontextprotocol/inspector band-mcp

🔨 Available Tools

Tool definitions live in band-sdk (see band.runtime.tools.iter_tool_definitions). The MCP server enumerates them at startup based on --scope and --tools. Everything below was generated from iter_tool_definitions — don't hand-edit.

Tool counts:

🤖 Agent tools (--scope agent)

For AI agents authenticated with an agent API key (band_a_*). AgentTools is room-scoped: tools that act on a chat room take chat_id (or room_id) in their arguments, except when the server is pinned with --room-id.

Baseline (always on):

Contacts — opt-in via --tools contacts:

Memory — opt-in via --tools memory:

👤 Human tools (--scope human)

For users authenticated with a user API key (band_u_*).

Baseline (always on):

Contacts — opt-in via --tools contacts:

Memory — opt-in via --tools memory:

💡 Usage Examples

Agent Framework Examples

We provide complete examples showing how to integrate Band MCP tools with popular agent frameworks. All examples use langchain-mcp-adapters to load the MCP tools.

Prerequisites for all examples:

• OpenAI API key (for the LLM)
• Band API key

Installation Options:

# Install dependencies for ALL examples
uv sync --extra examples

# OR install dependencies for specific frameworks:

# LangGraph only
uv sync --extra langgraph

# LangChain only
uv sync --extra langchain

LangGraph Agent

Uses LangGraph's StateGraph for building agents with MCP tools.

# Set your API keys
export OPENAI_API_KEY="sk-..."
export BAND_API_KEY="band_..."

# Run the interactive agent
uv run examples/langgraph_agent.py

What it does:

• Loads the Band MCP tools advertised by the server (see the tool counts table above)
• Creates an interactive chat loop with a GPT-4o powered agent
• The agent can manage chats, send messages, manage participants, and more
• Type exit, quit, or q to exit

See examples/langgraph_agent.py for the complete implementation.

LangChain Agent

Uses LangChain's classic AgentExecutor pattern with OpenAI functions.

# Set your API keys
export OPENAI_API_KEY="sk-..."
export BAND_API_KEY="band_..."

# Run the interactive agent
uv run examples/langchain_agent.py

What it does:

• Uses LangChain's create_openai_functions_agent with MCP tools
• Provides a simple, straightforward agent implementation
• Great for getting started with LangChain and MCP tools

See examples/langchain_agent.py for the complete implementation.

⚙️ Configuration

Credentials and scope (new in v1.2.0)

band-mcp now takes explicit dual credentials and lets operators pick which scopes and tool groups to serve:

# One credential per scope
export BAND_USER_KEY=band_u_your_user_key
export BAND_AGENT_KEY=band_a_your_agent_key

# Serve both scopes in one process (default: agent only)
uv run band-mcp --scope agent,human

# Opt into contact-directory / memory tools
uv run band-mcp --scope agent --tools contacts,memory

# Pin the whole server to a single chat/room
uv run band-mcp --scope agent --room-id r_123

Resolution precedence per field: CLI flag > BAND_* env. The legacy BAND_API_KEY env is still honored as a fallback — see below.

Breaking change note for --tools. Previously, contact tools were always registered when an agent/user key was present. The new default is --tools [] (no optional groups). Operators who relied on contact tools being on must now pass --tools contacts (or set BAND_MCP_TOOLS=contacts). Memory tools remain opt-in via --tools memory.

Unknown --scope / --tools values are logged at WARN with a "did you mean?" hint. Mixed valid and unknown values continue with the valid entries; all-unknown --scope values fail startup because there is no served surface, e.g.:

WARN  unknown --tools value 'contact' — did you mean 'contacts'? ignoring.
WARN  unknown --scope value 'huamn' — did you mean 'human'? ignoring.

Environment Variables

Legacy .env setups keep working unchanged:

# Legacy, still supported
BAND_API_KEY=your-api-key-here
BAND_BASE_URL=https://app.band.ai

When both a scope-specific key (BAND_USER_KEY / BAND_AGENT_KEY) and BAND_API_KEY are set, the scope-specific key wins for its scope. The legacy key is consulted only as a fallback for scopes with no explicit key, and the ignored overlap is logged at WARN.

Important: Never commit your .env file to version control. It's already in .gitignore.

🚨 Troubleshooting

Server Won't Start

# Check Python version (must be 3.11+)
python --version

# Verify the CLI is installed
band-mcp --help

# Try running with debug mode
BAND_LOG_LEVEL=debug band-mcp

Authentication Failures

• Verify your API key is correct and not expired
• Regenerate API key at app.band.ai/settings/api-keys
• Test API directly:

  curl -H "Authorization: Bearer $BAND_API_KEY" \
    https://app.band.ai/api/v1/health
  

AI Assistant Not Detecting Tools

1. Confirm band-mcp is on PATH: which band-mcp
2. Test server manually: BAND_API_KEY=... band-mcp
3. Restart your AI assistant completely
4. Check logs:

   # macOS
   tail -f ~/Library/Logs/Claude/mcp*.log
   

Common Error Solutions

💻 Development

Project Structure

band-mcp-server/
├── src/
│   └── band_mcp/              # Main package
│       ├── __init__.py            # Package initialization
│       ├── config.py              # CLI/env resolution, scope/tools parsing
│       ├── server.py              # MCP server entry point
│       ├── shared.py              # AppContext, HumanTools / AgentTools helpers
│       └── tools/
│           ├── __init__.py
│           └── registrar.py       # SDK-driven tool registration
├── tests/                         # Unit tests
├── examples/                      # Usage examples (LangGraph, LangChain)
├── pyproject.toml
├── .env.example
└── README.md

Tool implementations live in band-sdk (band.runtime.tools). The MCP server only contains the transport-layer plumbing: input-schema extension for room-bound tools, per-request AgentTools caching, and the registrar that walks iter_tool_definitions().

Setup Development Environment

# Clone the repository (with submodules for shared rules)
git clone --recurse-submodules https://github.com/thenvoi/thenvoi-mcp
cd thenvoi-mcp

# Copy environment template
cp .env.example .env  # then edit and set BAND_API_KEY

# Install with dev dependencies
uv sync --extra dev

# Install with ALL examples dependencies
uv sync --extra examples

# Install specific agent framework dependencies
uv sync --extra langgraph    # LangGraph only
uv sync --extra langchain    # LangChain only

# Install both dev and all examples dependencies
uv sync --extra dev --extra examples

# Install pre-commit hooks
uv run pre-commit install

Pre-Commit Hooks

This repository uses automated code quality tools:

• Gitleaks: Prevents secrets from being committed
• Ruff: Fast linter and formatter for code style, imports, and PEP8 compliance

The hooks will automatically check and format your code before each commit.

Local SDK Development

To develop against a local band-client-rest SDK instead of PyPI:

# 1. Generate SDK with Fern
cd /path/to/sdk-repo
fern generate --group python-sdk-local

# 2. Create package structure (Fern output needs wrapping)
mkdir -p sdk_package/band_rest
cp -r generated_sdk/* sdk_package/band_rest/

# 3. Create pyproject.toml for the package
cat > sdk_package/pyproject.toml << 'EOF'
[project]
name = "band-client-rest"
version = "0.0.1"
requires-python = ">=3.11"
dependencies = ["httpx>=0.25.0", "pydantic>=2.0.0"]

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"
EOF

# 4. Build wheel
cd sdk_package && uv build

# 5. Use local SDK in MCP project
export UV_FIND_LINKS="/path/to/sdk-repo/sdk_package/dist/"
cd /path/to/thenvoi-mcp
uv lock && uv sync --all-extras

After SDK changes:

# 1. Regenerate and rebuild wheel
cd /path/to/sdk-repo
fern generate --group python-sdk-local
rm -rf sdk_package/band_rest && mkdir -p sdk_package/band_rest
cp -r generated_sdk/* sdk_package/band_rest/
cd sdk_package && rm -rf dist && uv build

# 2. Clear uv cache and force reinstall
cd /path/to/thenvoi-mcp
uv cache clean --force band-client-rest
uv lock --upgrade-package band-client-rest
uv sync --all-extras

Important: You must clear the uv cache with uv cache clean --force band-client-rest before re-resolving. Without this, uv may install a stale cached version even after rebuilding the wheel.

Running Tests

# Run all tests with coverage
uv run pytest

# Verbose output
uv run pytest -v

# Run specific test file
uv run pytest tests/test_agents.py -v

# Generate HTML coverage report
uv run pytest --cov=src/band_mcp --cov-report=html

📚 Resources

• Model Context Protocol Documentation
• Band Platform
• uv Package Manager

Using Context7 MCP for Documentation

Context7 is an MCP server that provides up-to-date documentation for libraries and frameworks. It's highly recommended to use Context7 alongside Band MCP when developing—it helps your AI assistant fetch accurate, current documentation.

Adding Context7 to Your MCP Configuration

Add Context7 to your existing MCP configuration alongside Band:

{
  "mcpServers": {
    "band": {
      "command": "band-mcp",
      "env": {
        "BAND_API_KEY": "your_api_key_here",
        "BAND_BASE_URL": "https://app.band.ai"
      }
    },
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp@latest"]
    }
  }
}

Note: Context7 requires Node.js and npm/npx to be installed on your system.

How to Use Context7

Once configured, you can ask your AI assistant to fetch documentation:

• "Look up the Band REST API documentation with Context7"

Context7 will retrieve current documentation directly from official sources, ensuring your AI assistant has accurate information when helping you code.

📄 License

MIT