Skip to content

See how your app actually fits together. An interactive, plain-English map of your codebase, drawn by Claude or Codex.

9.2 KB
Raw Blame

How it works

Floorplan splits the work between scripts and the AI you run it in, Claude Code or Codex. The scripts do everything mechanical: finding files, parsing code, matching requests to endpoints, laying out the map. They're plain Node.js with no dependencies, and they cost no tokens. The AI does what only a reader of the code can: deciding what the parts of your app are, and explaining them plainly.

flowchart TD
    status["status: is there a map already?"] -->|no| scan
    status -->|yes| refresh["refresh: re-scan, carry over what didn't change"]
    scan["scan: repos, files, imports, endpoints, requests"] --> arch["The AI writes architecture.json<br/>the parts of the app, outside services, flows, tour"]
    arch --> build["build: assign files, connect everything,<br/>Health, screens, the map"]
    build --> describe["The AI names and describes every file and endpoint"]
    describe --> recent["The AI marks recent work"]
    refresh --> describe
    recent --> open["open: serve the map and show it"]

1. Scan

scan.mjs finds every git repository under the folder you point it at (one repo, a monorepo, or several repos side by side), skipping duplicate worktrees. It lists files with git, so your .gitignore is respected, then extracts from each file:

  • imports, resolved to real files: TypeScript path aliases, Python packages, Go modules, Rust crates, JVM packages;
  • exported definitions and their doc comments;
  • endpoints the server defines;
  • web requests the code makes;
  • content fingerprints, so later updates can tell what changed.

Supported languages: TypeScript and JavaScript (including Vue, Svelte and Astro files), Python, Go, Rust, Java, Kotlin, Scala and Groovy, with lighter support for others.

2. Connecting requests to endpoints

The most useful lines on the map connect a screen's web requests to the server code that answers them, often across repos. Floorplan:

  • Finds endpoints in Express, Fastify (plugin prefixes included), Koa, Hono, Next.js (route files and API routes), Firebase Cloud Functions, Flask, FastAPI, Django, Go (net/http, Gin, Echo, chi, Fiber), Actix, Axum and Spring, and in route tables keyed by method and path ("GET /api/orders": listOrders) in JavaScript and Python.
  • Resolves prefixes: Flask blueprints registered with a url_prefix, and Express routers mounted with app.use('/api/users', usersRouter), including routers mounted inside other routers and mounts with middleware in front.
  • Finds requests made with fetch, axios, ky, requests and similar, including:
    • addresses built in a variable first: const url = …; api.get(url);
    • template strings with nested parts;
    • small helpers in the same file that add a prefix, like call('/definitions').
  • Firebase Cloud Functions are called by name rather than by address. A callable function (onCall) shows up as CALL /approveItem under the name your functions entry file exports it as, and the website's httpsCallable(functions, 'approveItem') calls reach it, including through a helper of your own that wraps httpsCallable. An onRequest function is an ordinary endpoint at /<name>, and an Express app passed to it keeps its own routes under that name. Scheduled and other background functions (Firestore, Storage, Pub/Sub) are listed for the AI to place on the map.
  • Matches the two. When a request only spells out part of the address, because the base URL lives in settings, Floorplan learns each client's base path and prefers matches that fit it.

It's pattern matching, not a compiler, so it catches most connections but not all. If an important link is missing, ask for it to be added to the floorplan, and it will record it as an explicit connection.

3. The AI describes the system

Claude or Codex reads the scan summary, READMEs, setup files and entry points, then writes architecture.json:

  • the main parts of your app, each with a plain summary;
  • the outside services it really talks to (databases, payment providers, email);
  • connections the scan can't see, like queues, scheduled jobs and webhooks;
  • a few guided flows and a "Start here" tour.

The build then assigns every file to the deepest part whose patterns match, and reports anything left over.

Next, describe.mjs writes a worksheet for each part. A worksheet is a digest of every file and endpoint: what it defines, its doc comments, what it uses and what uses it. The AI fills in a name and description for each. On large projects, Claude Code splits the worksheets across parallel helpers.

By default everything is written for the app's owner, not an engineer: what each thing does for the business, in everyday words. Ask for a developer-facing map and it switches to technical style.

4. Build

build.mjs validates everything the AI wrote and reports problems by name ("did you mean …"). It then:

  • connects the parts, using imports, matched requests and the AI's explicit connections;
  • runs the Health check;
  • detects screens: Next.js, React Router, Vue, SvelteKit, Remix, Nuxt, Astro, Expo, Angular, TanStack;
  • writes graph.json and the self-contained index.html, noting whether Claude Code or Codex made it, so the map's Ask bar and fix prompts name the right one.

The map lays itself out in your browser with ELK, so the build itself is fast.

5. Recent work

After every make or update, the AI records what was just worked on in recent-work.json. recent.mjs lists the facts: what changed since the last map, uncommitted changes, and recent commits, each matched to its card. The AI adds the why: what you asked for, and how each change serves it. The previous highlights are replaced every time.

6. Showing it

open.mjs starts the map's small local server in the background and opens your browser. If the server is already running for this map, it's reused, and any open tab reloads itself in place instead of a new one opening.

Updates

refresh.mjs re-scans and compares against the last build, using each file's fingerprint:

  • exact content, which spots edits, renames and moves;
  • signature: its exports, endpoints and doc comments. Edits inside function bodies don't count.

It then:

  • carries descriptions over to moved and renamed files and endpoints;
  • drops entries for deleted files;
  • flags entries whose signature changed, for review.

Then it rebuilds. The AI only has to describe what's new or flagged, and check the structure if the change was large. That's why updates are quick.

Blueprints

Everything above is how Floorplan maps existing code, its main job. A blueprint is the other way in: a map of something that isn't built yet. The AI reads a PRD or a description and writes floorplan/plan.json: every file and endpoint the system would need, which screen calls which endpoint, which parts use outside services and the database, and who may use each endpoint. It also lists the stack it chose, its assumptions and the questions the PRD leaves open.

plan.mjs turns the plan into the same scan scan.mjs makes from code, so the build, Screens, flows, the tour and Ask all work the same way. The differences:

  • cards have dashed outlines and the board a Blueprint label;
  • Health becomes a design review of the plan: endpoints anyone could use to change data, endpoints with no decision on who may use them, open questions, and pieces nothing connects to;
  • every part has Build this with Claude (or Codex): a request to build that part as planned, listing its files, endpoints and what it relies on;
  • Ask answers from the plan and the PRD, since there's no code to read;
  • the Ask bar's Change switch takes requests like "also plan out refunds": the agent returns the updated plan, the map's server checks it with plan.mjs and build.mjs (sending problems back once), saves it and redraws the map with the changes highlighted, and Undo puts the previous plan back. Each design review item's Resolve on the map fills in that request for you: an open question just needs your answer. The agent stays read-only: only the server writes, and only plan.json, architecture.json and recent-work.json.

Updating a blueprint redraws it from plan.json. Once the code exists, refresh.mjs --from-code maps the real code instead, and plan.json stays.

Asking questions

The map's server answers each question with the AI on your computer, in your project:

  • Claude Code: it runs claude -p with read-only tools only: reading and searching files, and floorplan/query.mjs, which looks things up in the map.
  • Codex: it runs codex exec in Codex's read-only sandbox, with the same lookups. These runs aren't saved to your Codex history, so a follow-up brings the last few questions and answers along.

It uses whichever made the map, and the other one if that isn't installed. Either way it answers with text plus the cards to show, so answers can cite real code.

In the Claude app, the map asks Claude through the app instead, with the same lookups but no access to your source.

The same lookups help ordinary Claude Code sessions. If you say yes after your first map, the short section Floorplan adds to CLAUDE.local.md tells them to check the map (query.mjs find, show, path, overview) before searching the code from scratch.