2,625 stars in eight days: Floorplan, a Claude Code and Codex plugin created on October 27, 2026, has the agent read your code and draw an interactive map of the app, with every part, file and HTTP endpoint named in plain English and linked by its imports and calls. It is at version 0.6.3, MIT-licensed with zero dependencies, and its security documentation is more thorough than that of many plugins years older.
What it does
The floorplan skill kicks in when you ask to visualize, map or document a codebase, understand how the frontend talks to the backend, or update an existing map. The split of work is the interesting part: Node scripts handle the mechanics (finding repos, parsing imports and routes, matching each frontend HTTP call to the endpoint that serves it, even across separate repos, and laying out the graph), while the agent does what only a reader can: decide what each piece is and describe it in one line.
The output is floorplan/index.html at the project root, a map with several zoom levels (system, part, file group, file and endpoints). On top of that you get a Health check that looks for keys in code, open Supabase tables, permissive Firebase rules, unverified Stripe webhooks and unauthenticated endpoints; a Recent work view that highlights what the agent just touched and why; and an Ask bar for questions about the map ("what breaks if I change the users table?"). It detects routes in Express, Fastify, Next.js, Flask, FastAPI, Django, Spring, Gin, chi, Axum and Firebase Cloud Functions, among others.
Once a map exists, refresh.mjs compares content fingerprints with the last build and only asks the agent to describe what is new or whose interface changed, so updates are cheap in tokens. Past roughly 150 files, the skill splits the descriptions across parallel subagents.
Installing it
claude plugin marketplace add haydenschmitty/floorplan
claude plugin install floorplan@floorplanIn Codex, run codex plugin marketplace add haydenschmitty/floorplan and codex plugin add floorplan@floorplan. It needs Node 18 or later and runs on your existing Claude or ChatGPT plan, with no API keys. Then just ask "Make me a floorplan of this app" in a new session. There is a demo you can browse without installing anything.
Risks and limits
The SKILL.md lists eight commands in allowed-tools that run without asking while the skill is active, all of them node on its own scripts (build.mjs, describe.mjs, open.mjs, plan.mjs, recent.mjs, refresh.mjs, scan.mjs and status.mjs); anything else goes through your usual permissions. According to its "Is it safe to install?" page, it has no hooks, MCP servers, dependencies or install scripts, uses no eval, and a supply-chain test checks all of that on every change. The programs it starts are read-only git, node, the browser opener, and claude -p or codex exec --sandbox read-only when you ask the map a question.
What it does do, and you should know:
- A local server (
floorplan/serve.mjs) listens on127.0.0.1behind a token that only the map page receives. - One outbound request, opt-in: the Floorplan Cloud waitlist, if you type your email into the map.
- The Health check base64-decodes tokens it finds to classify them; it never stores or shows the value.
- It vendors a 1.6 MB minified file, the elkjs 0.9.3 layout engine, with its SHA-256 published so you can compare it with the npm package.
- If you agree, it adds a block to
CLAUDE.local.mdso future sessions query the map; that changes how those sessions work, which is why it offers rather than does it.
The underlying limit is the one every plugin has: an update brings whatever the latest release contains. The author publishes CHECKSUMS.txt with each version and a script to compare your installed copy; on a team, pinning the version and verifying it before updating is the sensible move. And the code the agent reads to describe it goes to Anthropic or OpenAI under your account, as with any other task.
Versus what was already there
The README compares itself with Archify, which draws polished diagrams of ten types for docs and talks, and with Understand Anything, which builds a function-level knowledge graph for studying large codebases. Floorplan sits in between: less detail than a function graph, but with the real frontend-to-backend connections, the Health check and the recent-changes view. It is written for people who build with AI and do not read the code; for an engineer, the value is in the developer view, the cross-repo map, and using it to review what the agent touched after a long task. Compared with hand-maintaining a Mermaid diagram, it regenerates itself; the catch is that the descriptions come from the model and need reviewing.
Sources
- haydenschmitty/floorplan (GitHub)
- floorplan SKILL.md (GitHub)
- Is it safe to install? (GitHub)
- Floorplan demo (Floorplan)