架构上下文增强MCP服务器

Enrich MCP

An MCP (Model Context Protocol) server that acts as an AI Architect enrichment layer on top of code graph data. It takes raw output from a code graph tool (e.g. codegraph), analyzes it using an AI model, and returns a structured architectural breakdown with a concise, token-efficient summary — so your AI coding assistant can write accurate code without hallucinating.

What does it do?

Exposes a single MCP tool:

enrich_architectural_context(raw_codegraph_data: str, target_focus: str) -> str

Pass in raw codegraph output and a focus (function name, file, or feature). The tool sends it to your chosen AI provider and returns two XML-tagged sections:

• <pre_thinking> — architectural breakdown: data flow, state mutations, key dependencies, and "gotchas".
• <distilled_context> — a concise markdown summary with exact signatures, DTOs, and caller contexts ready for use by a coding assistant.

Supports two AI providers: Anthropic (default) and GitHub Copilot.

Getting Started

1. Install Requirements

pip install -r requirements.txt

2. Configure Environment Variables

Export these variables in your shell:

Default values (hardcoded in enrich_mcp.py):

export ENRICHER_PROVIDER=anthropic
export MAX_OUTPUT_TOKENS=4096
export COPILOT_MODEL=claude-sonnet-4.5
export ANTHROPIC_MODEL=claude-haiku-4-5

Starting the MCP Server

python enrich_mcp.py

Example of Github Copilot CLI MCP configuration ~/.copilot/mcp-config.json:

{
  "mcpServers": {
    "enricher": {
      "type": "local",
      "command": "/path/to/enrich_mcp/.venv/bin/python",
      "args": ["/path/to/enrich_mcp/enrich_mcp.py"],
      "tools": ["*"],
      "env": {
        "ENRICHER_PROVIDER": "copilot", // or "anthropic"
        "MAX_OUTPUT_TOKENS": "4096",
        "COPILOT_MODEL": "claude-sonnet-4.5",
        "ANTHROPIC_MODEL": "claude-haiku-4-5",
        "ANTHROPIC_API_KEY": "sk-ant-abc123..."
      }
    }
  }
}

The server communicates over stdio (standard MCP transport). Register it in your MCP host (e.g. Claude Desktop, VS Code) pointing to this command.

To test the server end-to-end, put sample payload into sample/mcp_input_sample.json:

{
  "raw_codegraph_data": "function foo(...args) { ... }",
  "target_focus": "Callers of foo() and the data they pass in"
}

And then, run the test script:

cd /path/to/enrich_mcp
python -m scripts.enrich_test

If the server is running and properly configured, you should see the enriched architectural context printed in the console.

GitHub Copilot Provider

Login / Obtain a Token

When ENRICHER_PROVIDER=copilot, the server needs a GitHub OAuth token. Run the login script once to authenticate via GitHub Device Flow:

cd /path/to/enrich_mcp
python -m scripts.copilot_login

Follow the printed instructions — visit the URL and enter the device code in your browser. The token is saved to ~/.config/enrich_mcp/oauth_token.json and reused on subsequent runs.

Note: If you already have GitHub Copilot authenticated in VS Code, the token at ~/.config/github-copilot/hosts.json is picked up automatically — no login step needed.

Anthropic Provider

Set ENRICHER_PROVIDER=anthropic (the default) and provide your Anthropic API key via the environment variable:

export ANTHROPIC_API_KEY=sk-ant-...

The AsyncAnthropic client picks up ANTHROPIC_API_KEY automatically.