markdown-graph

Interactive force-directed knowledge graph of a repository's Markdown corpus. Every tracked .md file becomes a node, internal cross-references become edges, and each node renders as a pie chart sliced by its top 3 git contributors. Folder regions render as translucent hulls under the nodes so codebase structure is visible at a glance.

Built for "LLM context visualization" — seeing what a model would traverse when reading your docs — but works equally well as a generic "what's connected to what" tool for any repo with non-trivial Markdown.

Quick start

cd path/to/your/repo
npx -y markdown-graph
# → opens http://127.0.0.1:8765/ in your browser

That's the whole flow. The tool finds your tracked .md files via git ls-files, builds the graph, vendors force-graph from the local node_modules (no CDN), and serves the viewer.

Compare with other repos

Pass --repos to combine the current directory with one or more remote GitHub repos:

npx -y markdown-graph --repos vercel/next.js,facebook/react
npx -y markdown-graph --repos .,vercel/next.js

Each owner/name is sparse-cloned (--filter=blob:none + sparse-checkout for *.md only) into $TMPDIR/markdown-graph-cache/, so re-runs reuse the clone. The viewer's dropdown lets you switch between datasets without reloading.

Options

-r, --repos <list>   comma-separated: "." | "./path" | "/abs/path" | "owner/name"
-p, --port <n>       port (default 8765)
    --no-open        don't auto-open the browser
-b, --build <dir>    build a static site to <dir> and exit (no server)

--build is what the CI workflow uses to deploy the example site to GitHub Pages. The directory is self-contained — index.html, vendor/force-graph.min.js, datasets.json, and data/<name>.json — so any static host serves it.

What gets parsed

Every tracked .md file is a node. Edges come from four kinds of references:

Syntax	Example
Standard markdown link	[text](path/to/file.md)
Wiki-link	[[basename]]
Claude Code @-include	@path/to/file.md
Backtick relative path	`path/to/file.md` (requires a /)

Bare basenames inside backticks (`CLAUDE.md`, `README.md`) are intentionally skipped — they're usually nouns, not references, and would resolve ambiguously when the same basename exists in many subdirectories.

CHANGELOG.md files are excluded from the corpus. Contributor name variants (case, whitespace) are collapsed via a normalization pass and %aE/%aN are mailmap-aware, so a .mailmap in the source repo merges transliterations and nicknames.

Library use

const {generate} = require('markdown-graph/lib/generator');
const {runRepos, parseRepoList} = require('markdown-graph/lib/runner');

// Produce just the JSON:
await generate({root: '.', out: 'graph.json'});

// Generate + serve programmatically:
await runRepos({repos: parseRepoList('.,vercel/next.js'), port: 3000});

Layout

bin/markdown-graph.js   CLI (commander)
lib/generator.js        pure  generate({root, out, concurrency})
lib/runner.js           pure  runRepos({repos, port, openBrowser}) + sparse-clone
web/index.html          Viewer (vanilla JS + force-graph)

Dependencies

Four small runtime deps, vendored or required:

	Why
commander	CLI parsing
sirv	static HTTP serving
open	cross-platform browser launch
force-graph	the renderer; copied into the served temp dir at runtime

git (and optionally gh for private repos) must be on PATH. Node 18+.

License

MIT