Quasar Graph

Website · Demo · Docs · Install

A Claude Code plugin that turns a directory of projects into a living architecture graph — project and service nodes plus their databases, caches, queues, gateways, and third-party services, connected by dependency and ingress-route edges. The graph lives at quasar-graph.com: explore it in the web client, inspect any node, arrange the layout, share a deep link with a teammate.

Once a graph exists, the plugin uses it as a map of your codebase: answer cross-project questions, research code across many repositories, bind your current working project to its graph node, or clone an entire graph's projects onto a new machine.

Demo

From a scan to a living architecture graph in 33 seconds — watch the demo.

Install

/plugin marketplace add o-samoilov/quasar-graph
/plugin install quasar-graph@quasar-graph

Requires Claude Code and Node.js ≥ 20 for the bundled MCP server. The first backend call opens your browser to sign in with your quasar-graph.com account. Every skill that talks to the backend needs the server; only scan can run without it, keeping its result on disk for a later upload. Details: installation docs.

Skills

Command	What it does
/quasar-graph:scan	Scan a directory of projects and build/update a graph: catalog → deep analysis (parallel sub-agents) → resource aggregation → pre-upload review → upload.
/quasar-graph:context	Bind the current working project to its graph node (matched by git remote), then answer structural questions from the graph — who calls this service, what does it depend on, what routes traffic to it — and, with confirmation, look into dependent projects' code.
/quasar-graph:research	Graph-guided, read-only code search across projects: scope relevant nodes from the graph, auto-clone missing repos, dispatch per-project research sub-agents, and synthesize an answer with file:line references.
/quasar-graph:clone	Clone (and optionally install) all projects of a graph onto the local machine, with a deterministic clone plan (target paths, ssh/https, monorepo dedup).
/quasar-graph:edit	Edit an existing graph without re-scanning: pull the full snapshot, change positions, descriptions, agent context, nodes and edges in conversation, review a change summary, and push it back (full replace).

Three internal skills (analyze-project, research-project, install-project) are dispatched by the orchestrators via sub-agents and are not meant to be invoked directly.

How it works

The scanner is agent skills (markdown instructions — no application code): Claude Code reads project manifests and config files, writes structured markdown per project, and a small bundled Node.js MCP server (mcp/) transforms the result into nodes/edges and syncs it with the backend.

• Config only, never source code. The analyzer reads manifests, .env* files (names only, never values), compose files, ORM configs, and the root README.md. Better to miss a dependency than to fabricate one.
• Two edge layers. Logical dependency edges (who calls / needs whom) are the source of truth; the gateway/ingress routing path (Traefik, nginx, …) is a separate, additive routes overlay.
• Deterministic ids. Nodes and edges get UUIDv5 ids, so re-scans reuse the same ids and graph-owned fields (node positions, per-node agent instructions) survive.
• Human-in-the-loop. The scan gates on your approval before deep analysis and again before upload, with an automated pre-upload review that flags likely mistakes.

Full scan flow, phase by phase: scan docs.

The quasar-graph backend and web client are separate applications and are not part of this repository — this repo implements the scanner skills plus the MCP client that syncs with them.

Development

The skills are markdown and have no build step. The MCP server does:

cd mcp
npm install
npm test        # Node's built-in test runner, suites under mcp/test/
npm run build   # regenerates the committed bundle mcp/dist/index.js

The plugin runs the bundled mcp/dist/index.js, so rebuild after editing mcp/src/ or mcp/index.js. The server's tools, configuration, release flow, and backend contract are documented in mcp/README.md. See references/ for how the skills' harness-agnostic action vocabulary maps onto concrete harness tools (useful when porting the skills beyond Claude Code).

License

MIT