MCP Architect

🏛️ mcp-architect

Stop pasting your file tree into Claude. Give any AI assistant real architectural understanding of a codebase — local, private, zero‑config.

AI coding assistants are great at files but blind to architecture. Every session you re‑explain the structure, paste the file tree, and hope it guesses your module boundaries right. mcp-architect is an MCP server that hands your assistant a structured map of any codebase — tech stack, dependency graph, hotspots, and module summaries — computed 100% locally with no API keys and no model required.

It works with Claude Desktop, Cursor, Windsurf, Cline, or any MCP client.

Why

Everything runs on your machine. Your code never leaves it.

Quickstart

Install

pip install mcp-architect

…or skip the install entirely and let your MCP client fetch it on demand with uvx (shown below).

1. Add it to your MCP client

Claude Desktop — edit claude_desktop_config.json:

{
  "mcpServers": {
    "architect": {
      "command": "uvx",
      "args": ["mcp-architect"]
    }
  }
}

Prefer pip? pip install mcp-architect. Or run the latest straight from source:

{ "mcpServers": { "architect": {
    "command": "uvx",
    "args": ["--from", "git+https://github.com/kannajune/mcp-architect", "mcp-architect"]
} } }

Restart your client. That's it — no keys, no model download.

2. Ask your assistant

"Use the architect tools to give me an overview of ~/code/my-app, then show me its dependency graph and the highest‑risk files."

What you get

# Architecture Overview — my-app

**151 files · 17,368 lines of code**

## Languages
- **Python** — 93 files, 13,683 LOC
- **TypeScript** — 23 files, 3,120 LOC

## Frameworks / key libraries
- FastAPI
- React
- Tailwind CSS

## Entry points
- main.py

# Dependency Graph — my-app

**118 modules · 172 internal import edges**

## Most depended-upon (architectural hubs)
- `app.signals.signal_parser` — imported by 12 modules
- `app.core.integrations_registry` — imported by 11 modules

## Circular dependencies
✅ no circular dependencies found

Tools

Design principles

• Zero heavy dependencies. Pure Python standard library for all analysis (ast, os, re). The only runtime dep is the MCP SDK itself. Installs in seconds.
• Local & private. No network calls, no telemetry, no LLM. Your source never leaves your machine.
• Language‑aware. Full AST parsing for Python; import parsing for JavaScript/TypeScript; file/LOC stats for 25+ languages.
• Decoupled core. The analysis layer (mcp_architect.analysis) is importable and testable on its own — use it as a plain Python library too.

from mcp_architect.analysis import get_overview, get_dependency_graph
print(get_overview("~/code/my-app")["frameworks"])

The dependency and complexity analysis is heuristic — designed to give an AI useful, fast situational awareness, not to replace a full static analyzer.

Pin to one project (optional)

Set MCP_ARCHITECT_ROOT so tools default to a fixed repo and you can omit paths:

{ "mcpServers": { "architect": {
    "command": "uvx", "args": ["mcp-architect"],
    "env": { "MCP_ARCHITECT_ROOT": "/Users/you/code/my-app" }
} } }

How it compares

mcp-architect isn't a semantic search engine or a context packer — it's a structural lens any AI assistant can call on demand. It's designed to complement the tools below, not replace them:

In one line: Cursor and Serena help the AI read your code; mcp-architect helps it understand the architecture — locally, in any client. They stack well together.

Roadmap

• [ ] Layered‑architecture / boundary‑violation detection
• [ ] Go, Rust & Java import graphs
• [ ] Optional local‑LLM (Ollama) narrative summaries
• [ ] compare tool for before/after architecture diffs

Contributions welcome — see CONTRIBUTING.

Contributing

PRs and issues welcome! Run the tests with:

pip install -e ".[dev]"
pytest

License

MIT © Kannan Dharmalingam