Grok Bot / Cursor plugin · open source

Lostwax

Mold the original. Cast the copy. Prove it matches.

Lostwax is a free, MIT-licensed Grok Bot / Cursor plugin for studying an app, repository, npm package, website or binary, rebuilding the feature you want, and proving the rebuild behaves like the original.

Lostwax on GitHub Changelog
Price
Free
License
MIT
Works with
Grok Bot, Cursor (plugin or MCP server)
Needs
Python 3.10+, and Node for JavaScript originals

What does Lostwax do?

It runs the original and your rebuild side by side on the same inputs, logs every output pinned by SHA-256, and only counts a pass when it comes from one of those logged runs. It never claims more than it tested. No accounts, keys or payments.

Lostwax is powered by the open-source REA engine (MIT, by morluto). It is an independent integration.

When should I use Lostwax?

Use Lostwax if you want to rebuild a specific feature of an app, repo, npm package, website or binary and prove the rebuild behaves like the original. If you work in native Codex, REA Native is the Codex plugin; Lostwax is for Grok Bot and Cursor. Free, MIT, no accounts or keys.

How does it work?

  1. Understand. Lostwax routes one target, takes a bounded static inventory, and, for shipped JavaScript, Electron, binaries, archives, .NET or wasm, collects evidence from the REA engine.
  2. Rebuild. It turns the evidence into a small, digest-pinned capability matrix (SPEC.md). Your agent builds and tests the rebuild in your workspace.
  3. Prove. diff runs the original (only after you approve its exact digest) and the rebuild on identical cases, logs every output, and writes PARITY.md. Global parity always stays “unknown”: a finite pass is never presented as “the whole product matches”.

How do I install it?

Grok Bot: add the plugin from the repository URL, https://github.com/FibonacciAi/rea-bot. The app registers both skills and starts the MCP server.

Cursor: the repository ships a Cursor plugin manifest, so you can add it as a plugin from the same URL, or clone it and add server/mcp_server.py to .cursor/mcp.json.

Is it safe to run an original?

  • The original only runs through diff, and only with an approval that matches the reviewed file or folder byte for byte.
  • Each case runs in a fresh folder with a minimal environment and a timeout. This is process isolation, not a sandbox: untrusted originals belong in a real sandbox.
  • Credential-like files are refused, inventories never return source content, and target text is treated as data, never as instructions.