repo-cartographer

Understand any codebase in 60 seconds. An MCP server that turns any repository into an architecture diagram.

Point your LLM at a folder and get back an architecture map: languages, frameworks, entry points, modules, and an import graph — rendered as a Mermaid diagram you can drop into a PR, a doc, or an onboarding guide.

It is model-agnostic. It speaks the Model Context Protocol over stdio, so it works with Claude Code, Claude Desktop, Cursor, Cline, Copilot, or the OpenAI Agents SDK — no Claude-specific dependency.

Why it's different

The server extracts hard facts. The model does the reasoning.

repo-cartographer never tries to "understand" your code semantically. It parses deterministic facts — file tree, import edges, manifests, framework detection, entry points — and hands your LLM structured JSON plus a draft diagram. Your model turns those facts into the final narrative and a refined diagram.

That split keeps the server small, fast, testable, and portable — and it means the diagram is grounded in what's actually in the repo, not hallucinated.

Example output

Running generate_diagram against this repo produces (a draft the model then refines):

flowchart TD
    n0["src · 2 files"]
    n1["src/lib · library code · 9 files"]
    n2["src/resources · resource handlers · 1 file"]
    n3["src/tools · tool implementations · 4 files"]
    n0 --> n1
    n0 --> n2
    n0 --> n3
    n1 --> n0
    n3 --> n0
    n3 --> n1

A self-contained, shareable HTML render is committed under examples/ (both a high-level and a file-level detail view).

Install & run

Requires Node.js 18+.

# Run directly (no install)
npx -y repo-cartographer

# …or from source
git clone https://github.com/builditwithgk/repo-cartographer
cd repo-cartographer
npm install
npm run build
node dist/index.js

The server communicates over stdio; AI clients launch it that way. You can also use it directly from a terminal — see below.

Command line (no AI needed)

The same binary is dual-mode: with no arguments it's the MCP server; with a subcommand it's a plain CLI for humans and CI.

# Draw a diagram — path in, architecture.html out
npx -y repo-cartographer map ./my-project
npx -y repo-cartographer map ./my-project --level detail -o docs/architecture
npx -y repo-cartographer map ./my-project --format dot   # Graphviz DOT instead of Mermaid

# Enforce architecture rules (exits 1 on an error-level violation — use it in CI)
npx -y repo-cartographer check ./my-project --config .cartographer.yml

Zip-friendly: if you point it at an extracted "Download ZIP" folder (repo-main/ wrapper and all), it detects the wrapper and maps the real repo root — noted in the output, never silent.

Two output notations, one strategy: Mermaid because that's where people read it (GitHub renders it natively in PR comments and READMEs), DOT because that's what their tools eat (pipe it into Graphviz, Backstage, or anything else: dot -Tsvg architecture.dot). Both come with the same shareable HTML page and role-colored modules. When the repo has a .cartographer.yml, its diagram: section supplies the defaults for --level, --format and -o; explicit flags always win.

check reads a rules file and flags forbidden cross-boundary imports and dependency cycles — deterministically, no LLM involved, so it's safe to gate a merge:

# .cartographer.yml
rules:
  forbidden:
    - from: "src/ui/**"
      to:   "src/db/**"
      reason: "UI must go through the service layer, not the DB directly."
  cycles: error        # error | warn | off

Adopting on an existing codebase? Record today's violations as an accepted baseline, so only new ones fail the build:

npx -y repo-cartographer check . --update-baseline   # writes .cartographer-baseline.json

Commit that file; later runs auto-detect it and pass unless a PR introduces a new violation.

Architecture governance in CI (GitHub Action)

A composite action (action.yml) runs check on every PR — failing the build on an error-level violation and posting a sticky comment with the diagram and any violations:

# .github/workflows/architecture.yml
on: pull_request
permissions: { contents: read, pull-requests: write }
jobs:
  architecture:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: builditwithgk/repo-cartographer@v1
        with: { path: ., config: .cartographer.yml }

This repo dogfoods it in .github/workflows/architecture.yml. Full design + phases: docs/github-action.md.

Use it with Claude Code

claude mcp add repo-cartographer -- npx -y repo-cartographer

Then ask: "Use repo-cartographer to map ./my-project and draw me an architecture diagram."

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "repo-cartographer": {
      "command": "npx",
      "args": ["-y", "repo-cartographer"]
    }
  }
}

Use it with the OpenAI Agents SDK

Same server, a different model — the whole point of MCP:

import asyncio
from agents import Agent, Runner
from agents.mcp import MCPServerStdio

async def main():
    async with MCPServerStdio(
        params={"command": "npx", "args": ["-y", "repo-cartographer"]},
    ) as cartographer:
        agent = Agent(
            name="Cartographer",
            instructions=(
                "Use the repo tools to gather facts, then refine the draft "
                "Mermaid diagram into a clean architecture map."
            ),
            mcp_servers=[cartographer],
        )
        result = await Runner.run(agent, "Map ./my-project and explain its architecture.")
        print(result.final_output)

asyncio.run(main())

Tools & resources

Most of the time you just want map_repo — "path in, diagram out." Reach for the four granular tools only when you want to compose the steps yourself (e.g. let the model refine the diagram source between generate_diagram and render_diagram).

How it stays fast on big repos

• Languages: JavaScript/TypeScript and Python (v1).
• Skips node_modules, .git, dist, build, venv, __pycache__, vendor, and other build/dependency/cache directories (plus all hidden dirs).
• Caps the number of files scanned and bytes read per file; collapses the import graph to directory level past a threshold. Anything capped is reported in the output — never dropped silently.
• No network calls at scan time. (The rendered HTML pulls Mermaid from a CDN only when you open it in a browser.)
• Deterministic: output is sorted and stable, so diagrams don't churn between runs.

Development

npm run dev        # run from source with tsx
npm run build      # type-check + emit to dist/
npm test           # unit tests (node:test, no build step needed)
npm run typecheck  # type-check src/ and test/ together, no emit

Tests cover the deterministic core — import resolution (JS/TS + Python), module collapsing, cycle detection, rule globs, baseline diffing, Mermaid rendering and the check end-to-end path. They run against src/ via tsx, so there is no build step and no test framework dependency.

Author

Built by Gopi K Aitham (builditwithgk) — see about://author, or scaleup-solutions.in. Available for freelance and contract work on AI, agent, and MCP tooling.

License

MIT