The Cursor team shipped Grok Bot 0.18.0 with runtime source maps enabled. Bennett used them to reconstruct readable runtime source from the packaged macOS app.
The result offers a rare look inside a production coding agent. The repository includes code around system prompt assembly, tool definitions, inference routing, the Electron host, the coordinator, and local execution.
What to Read
I would start with three layers:
- Prompt: how the agent assembles instructions and context.
- Tools: how it defines, registers, and approves actions.
- Routing: how model requests move through the inference layer.
The project also adds routing for Claude Code, Codex, and OpenRouter, plus an optional local Docker sandbox. Those are extensions by the researcher, not confirmed features from the original app.
One Important Caveat
This is not Cursor's original monorepo. The author calls it an unofficial, source-oriented reconstruction. The original frontend source and source maps were not included, so the project keeps the packaged renderer as its UI baseline.
I checked the latest Grok Bot build, and the runtime source maps are gone. Treat this as a snapshot of version 0.18.0 and a useful reference for studying production agent architecture.
Read the reconstructed source repository.
