Mattermost MCP Host

A Mattermost integration that connects to Model Context Protocol (MCP) servers, leveraging a LangGraph-based AI agent to provide an intelligent interface for interacting with users and executing tools directly within Mattermost.

Demo

1. Github Agent in support channel - searches the existing issues and PRs and creates a new issue if not found

2. Search internet and post to a channel using Mattermost-MCP-server

Scroll below for full demo in YouTube

Features

• 🤖 Langgraph Agent Integration: Uses a LangGraph agent to understand user requests and orchestrate responses.
• 🔌 MCP Server Integration: Connects to multiple MCP servers defined in mcp-servers.json.
• 🛠️ Dynamic Tool Loading: Automatically discovers tools from connected MCP servers and makes them available to the AI agent. Converts MCP tools to langchain structured tools.
• 💬 Thread-Aware Conversations: Maintains conversational context within Mattermost threads for coherent interactions.
• 🔄 Intelligent Tool Use: The AI agent can decide when to use available tools (including chaining multiple calls) to fulfill user requests.
• 🔍 MCP Capability Discovery: Allows users to list available servers, tools, resources, and prompts via direct commands.
• #️⃣ Direct Command Interface: Interact directly with MCP servers using a command prefix (default: #).

Overview

The integration works as follows:

1. Mattermost Connection (mattermost_client.py): Connects to the Mattermost server via API and WebSocket to listen for messages in a specified channel.
2. MCP Connections (mcp_client.py): Establishes connections (primarily stdio) to each MCP server defined in src/mattermost_mcp_host/mcp-servers.json. It discovers available tools on each server.
3. Agent Initialization (agent/llm_agent.py): A LangGraphAgent is created, configured with the chosen LLM provider and the dynamically loaded tools from all connected MCP servers.
4. Message Handling (main.py):
   • If a message starts with the command prefix (#), it's parsed as a direct command to list servers/tools or call a specific tool via the corresponding MCPClient.
   • Otherwise, the message (along with thread history) is passed to the LangGraphAgent.
5. Agent Execution: The agent processes the request, potentially calling one or more MCP tools via the MCPClient instances, and generates a response.
6. Response Delivery: The final response from the agent or command execution is posted back to the appropriate Mattermost channel/thread.

Setup

1. Clone the repository:

   git clone <repository-url>
   cd mattermost-mcp-host
   

2. Install:

   • Using uv (recommended):

     # Install uv if you don't have it yet
     # curl -LsSf https://astral.sh/uv/install.sh | sh 
     
     # Activate venv
     source .venv/bin/activate
     
     # Install the package with uv
     uv sync
     
     # To install dev dependencies
     uv sync --dev --all-extras
     

3. Configure Environment (.env file): Copy the .env.example and fill in the values or Create a .env file in the project root (or set environment variables):

   # Mattermost Details
   MATTERMOST_URL=http://your-mattermost-url
   MATTERMOST_TOKEN=your-bot-token # Needs permissions to post, read channel, etc.
   MATTERMOST_TEAM_NAME=your-team-name
   MATTERMOST_CHANNEL_NAME=your-channel-name # Channel for the bot to listen in
   # MATTERMOST_CHANNEL_ID= # Optional: Auto-detected if name is p
   

... View full README on GitHub