Understand any codebase in 60 seconds — an MCP server that turns any repo into an architecture diagram.
The server defines 5 tools with clear verb-first naming (map_repo, scan_repo, build_import_graph, generate_diagram, render_diagram). All tools have descriptions ranging from 194 - 380 characters, well within the 10 - 1024 baseline. Every tool has a properly typed input schema using Zod with descriptions on all parameters. The tool descriptions clearly state WHAT the tool does and WHEN to prefer it over alternatives (e.g., 'Prefer map_repo for the one-shot diagram flow; use build_import_graph when you need dependency edges'). However, output schemas are not documented, callers must infer structure from the code. Error handling is basic (try/catch with message() conversion); no guidance on recovery steps, retryability, or next actions. STDIO transport caps the protocol score at 50, but definition quality is solid for this domain.
Extract intra-repo import/require relationships for JavaScript/TypeScript and Python. Nodes are files (auto-collapsed to module/directory level for large repos); edges are directed import relationships. Returns JSON. Use this when you need the raw dependency edges — who imports whom — e.g. to reason about coupling, cycles, layering, or the blast radius of a change. Do NOT use it just to draw a picture: generate_diagram returns a ready-to-render draft, and map_repo does the whole flow in one call.
Combine scan + import facts into a DRAFT diagram string — Mermaid flowchart (default) or Graphviz DOT. level 'high' (default) shows modules/directories; 'detail' shows files grouped by module. This is a draft the calling model is expected to refine. Use this when you want to edit or refine the diagram source before rendering (pair with render_diagram to write the HTML afterwards). If no in-between editing is needed, prefer map_repo — it generates AND renders in one call.
Point at a folder and get a downloadable architecture diagram in one call: runs scan + import-graph + render internally and returns the HTML path, a facts summary, and the draft diagram source inline. By default writes architecture.html plus the raw source (.mermaid or .dot) into the repo; pass outPath to write elsewhere. Use the granular tools (scan_repo, build_import_graph, generate_diagram, render_diagram) only when you need to compose the steps yourself.
Write a self-contained, shareable HTML page (renders via CDN: Mermaid, or Viz for Graphviz DOT) plus the raw source file (.mermaid or .dot). Returns the absolute paths written. Use this as the final step after refining a draft from generate_diagram — or to render ANY Mermaid/DOT source, hand-written included. If you have not edited the draft, prefer map_repo, which renders it for you.
Output schemas are not documented. Callers cannot know what fields map_repo, scan_repo, build_import_graph, and generate_diagram return. The descriptions mention 'returns JSON' or 'returns HTML path and facts' but omit field names, types, and structure. This forces LLMs to infer output shape and increases risk of extraction errors.
Error handling provides no recovery guidance. The try/catch in server.ts simply calls message(e) and returns errResult(). No distinction between retryable errors (transient network failures), user-fixable errors ('path does not exist'), or fatal errors. LLMs cannot determine appropriate next actions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 65 | 2026-07-28+ | v2 |
Deterministically extract facts from a local repo: programming languages, frameworks/libraries, entry points, top-level modules (with role guesses), and a manifest summary. Returns JSON facts only — no interpretation. Use this when you need repo facts WITHOUT a diagram (e.g. to answer questions about the stack). Prefer map_repo for the one-shot diagram flow; use build_import_graph when you need dependency edges rather than an inventory.
map_repo and render_diagram perform state-altering operations (write files to disk, create .html and .mermaid/.dot files) but descriptions do not explicitly state this. 'Modifies state' and 'irreversible consequences' guidance are missing, which agents need to reason about retry safety and side effects.
No input validation or constraint documentation for path parameters. Descriptions say 'Absolute or relative path' but do not specify: must the path exist? Must it be readable? What happens if the path contains special characters or is a symlink? Missing constraint details force LLMs to pass invalid inputs and rely on error messages to discover boundaries.