mdgraph

Scan a repo's markdown into a graph, visualize it in the browser, and expose it to Claude (or any MCP host) as a navigation index.

One data model — { nodes, edges } — with three consumers:

              scan  (the only hard part)
                 │  emits { nodes, edges }
      ┌──────────┼───────────┐
      ▼          ▼           ▼
   graph.json  browser     MCP server
   (a file)   (viewer)   (agents read it)

Insights, reachability, and orphans are all derived from the model at read time. Nodes and edges stay dumb.

The model

node = { id, label, path, type, cluster, content }   // type: root | ai | doc
edge = { source, target, sourceLine, context }        // context = the line the link lives on

type is assigned by filename convention (README.md → root, CLAUDE.md/AGENTS.md/*.mdc → ai, else doc). Edges come from actual links in the text — standard [x](./y.md) and [[wikilinks]] — not folder membership, so a file in docs/ that links across the tree still connects across the tree. Folders only drive cluster (coloring/grouping).

Related MCP server: mcp-server-markdown

Install & run

No install needed — run it straight with npx:

npx @emmd474/mdgraph scan .            # print the graph as JSON
npx @emmd474/mdgraph view .            # open the interactive graph in a browser
npx @emmd474/mdgraph scan . --out graph.json

Or install globally to get a plain mdgraph command:

npm install -g @emmd474/mdgraph
mdgraph view ./my-repo
mdgraph scan ./my-repo --out graph.json

dir defaults to the current directory. view takes --port (default 4173).

Commands

Expose to Claude (MCP)

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "mdgraph": {
      "command": "npx",
      "args": ["-y", "@emmd474/mdgraph", "mcp", "/abs/path/to/your/repo"]
    }
  }
}

Claude Code:

claude mcp add mdgraph -- npx -y @emmd474/mdgraph mcp /abs/path/to/your/repo

Restart the host and Claude gets four tools:

The point: an agent calls get_project_map to orient, then get_context("README.md") to pull only the docs the task needs — instead of dumping every .md into context.

Publishing (for maintainers)

The name mdgraph is taken on npm, so this publishes scoped under your account:

# 1. set the scope to YOUR npm username (edit "name" in package.json if not @emmd474)
npm whoami          # confirm you're logged in; else: npm login
npm publish --access public

Scoped packages are private by default, so --access public is required the first time. To ship an update, bump the version and publish again:

npm version patch   # 0.1.0 -> 0.1.1, also creates a git tag
npm publish

npm pack --dry-run shows exactly what ships (currently 7 files, ~16 kB — no node_modules, no test repo, thanks to the files allowlist).

Try it on the example

git clone <this repo> && cd mdgraph && npm install
node cli.mjs view testrepo
node cli.mjs scan testrepo

What's real vs. next

Tested and working: the scanner (md + wikilink extraction, code-fence skipping, convention typing, orphan detection), the browser viewer, and all four MCP tools over a live stdio handshake.

Next, roughly in order:

• Harden link parsing — the scanner is line-based and skips fenced blocks, but an inline code span (`[x](y)`) can still yield a false link. Swap in a real markdown AST (remark + unist-util-visit) when you want precision; the emitted shape stays identical.

• Manual edges — let the viewer write related: into a file's frontmatter, read as a third edge source.

• Frontmatter tags as a cluster source alongside folders.

• --watch on view to push graph updates over server-sent events.

MIT licensed.