Home

Foldo

A Figma-style multiplayer review canvas for AI-generated code — comment, compare, and ship the changes Claude Code (and other agents) just wrote, without leaving the canvas. Now with User Tests: gather screen + voice evidence from real users and turn it into agent edits.

This wiki is a high-level orientation. The deep technical docs live in the repo under docs/.

What Foldo does

Writing variants with AI agents is cheap and fast — reviewing them isn't. Foldo folds parallel AI-generated work back into a single spatial canvas where humans direct, compare, and approve:

• Live app frames — iframes of the running app at a specific commit + reproducible state.
• Pinned comments → edits — turn any comment thread into a structured Claude Code prompt; the agent ships a new commit and a new frame appears on the canvas.
• Multiplayer — live cursors, presence avatars, follow-me viewports, over plain WebSockets.
• Capture anywhere — a Chrome extension freezes any deployed URL into a frame; a GitHub webhook auto-creates frames on push.

User Tests — real-user evidence on the canvas

Every reviewer on the canvas is a proxy for the real user. User Tests closes the loop that was missing — build → real users try it → evidence → fix:

1. Publish a link — pick an app URL, write task instructions + an optional questionnaire → a short foldo.dev/t/:token link.
2. Testers record — screen + voice (or voice-only) against the tasks; three delivery modes are auto-detected (embedded iframe, new-tab handoff, or a frozen DOM snapshot for local-only apps).
3. Results land as frames — each session becomes a test_session frame (scrubbable recording, per-task pass/skip stats, questionnaire answers, transcript, AI synthesis), clustered under a test_summary frame, streaming in live.
4. Feedback → edit — every AI-extracted issue carries a "Make this an edit" button into the existing dispatch pipeline.

See the User Tests design doc for the full rationale.

Architecture at a glance

Service	What it is
apps/web	The canvas + marketing site + the public /t/:token test runner
apps/server	Fastify REST + WebSocket gateway, Postgres-backed
apps/sample-app	Demo app iframed by the canvas
apps/mcp	In-directory MCP server — bridges Claude Code to the canvas
apps/shotter	Headless-Chromium screenshot fallback for un-iframeable URLs
packages/protocol	Shared TypeScript types + the wire protocol

Deep docs

• Architecture — components, data model, data flow, where the prototype ends.
• Protocol — REST + WebSocket + MCP wire protocol, frame kinds.
• User Tests — the unmoderated UX testing design.
• MCP · Extension · Deployment · Demo path

---

_{Built as a demo of what review tools look like when most code is written by agents. Branding inspired by the dachshund — small, focused, low to the ground.}