[RFC] Exposing AutoMaintainer as an MCP Server

[purvanshjoshi]

[RFC] Exposing AutoMaintainer as an MCP Server (npx @pxa-labs/automaintainer-mcp)

The Vision

The Model Context Protocol (MCP) is rapidly becoming the standard for Automation tool interoperability. By exposing the AutoMaintainer 5-agent LangGraph pipeline as an MCP server, we can decouple our autonomous capabilities from our proprietary web dashboard.

This enables any compliant Automation interface (Claude Desktop, Cursor, Devin, Windsurf) to dynamically discover, instantiate, and "hire" the AutoMaintainer crew as a subcontractor to execute complex repository-wide operations.

Proposed Architecture

The MCP Server will serve as a lightweight wrapper around the LangGraph execution engine. It will expose specific tool_calls that external AIs can invoke.

flowchart TD
    classDef external fill:#2d3748,stroke:#4a5568,stroke-width:2px,color:#fff,rx:8px
    classDef interface fill:#2b6cb0,stroke:#2c5282,stroke-width:2px,color:#fff,rx:8px
    classDef engine fill:#276749,stroke:#22543d,stroke-width:2px,color:#fff,rx:8px
    classDef transport fill:#744210,stroke:#5f370e,stroke-width:2px,color:#fff,rx:8px

    User(["Developer (User)"])
    
    subgraph HostEnvironment [Host Environment]
        HostAI["Host Automation Client\n(Claude / Cursor)"]:::external
    end
    
    subgraph MCPServer [AutoMaintainer MCP Server]
        Transport["JSON-RPC over STDIO\n(@modelcontextprotocol/sdk)"]:::transport
        Controller["Tool Invocation Controller\n(Node.js Wrapper)"]:::interface
        Engine["LangGraph 5-Agent Pipeline\n(Python Engine)"]:::engine
    end

    User -->|Prompts: 'Refactor Auth'| HostAI
    HostAI -->|Resolves tool: automaintainer_implement| Transport
    Transport <--> Controller
    Controller -->|Spawns Headless Subprocess| Engine
    Engine -->|Returns Execution Logs & Diff| Controller

Loading

Exposed Tools (Draft)

The MCP server will register the following tools with the host Automation:

1. automaintainer_analyze_repo(repo_path): Triggers the Architect and Visionary agents to scan the repository via GitNexus and return an architectural summary.
2. automaintainer_implement_feature(description, path): Triggers the Implementer and Reviewer agents to write, test, and verify code locally.
3. automaintainer_review_pr(pr_number): Triggers the Reviewer and Maintainer agents to analyze a specific PR and return actionable feedback.

Distribution Strategy

To eliminate friction, the server will be distributed via NPM. Users will configure their MCP clients (e.g., claude_desktop_config.json) with:

{
  "mcpServers": {
    "automaintainer": {
      "command": "npx",
      "args": ["-y", "@pxa-labs/automaintainer-mcp"]
    }
  }
}

Open Questions

1. How do we handle Python environment dependencies (LangGraph, FastAPI) when the user executes the server via npx? Should the NPX package handle bootstrapping a venv, or should we distribute this via pip instead?
2. How do we stream LangGraph execution logs back to the Host Automation in real-time using MCP's notifications protocol?

Please leave your thoughts and feedback below!