What's this about?
Days after a researcher alleged that xAI's coding tool Grok Build was secretly uploading users' code to xAI's cloud (yesterday's top story), xAI published the tool's entire source code on GitHub — 844,000 lines of Rust under the permissive Apache-2.0 license. Simon Willison spent the evening reading it and surfaced fun details, like a built-in Mermaid diagram renderer that draws charts in the terminal using Unicode box-art.
What exactly is "open-sourcing" a coding agent?
A coding agent is an AI tool that doesn't just autocomplete — it reads your repository, runs commands, edits files, and iterates until a task is done. The "harness" is all the non-AI plumbing around the model. Open-sourcing it means publishing that plumbing's source code so anyone can read exactly what it does — including whether it phones home with your data. It does *not* mean the Grok model itself is open; only the tool around it.
What's the community saying?
- "This is damage control, not generosity" — the dominant take: after being caught exfiltrating user data (env files, whole source trees — not just prompts), opening the code is the only trust-rebuilding move xAI had left.
- "Deleted the data" ≠ proof — commenters note xAI claims it will delete the exfiltrated data but has produced no third-party destruction certificate (Kroll, Epiq, etc.), which is the standard for actually trusting such claims.
- Credit where due — several say the harness itself is "butter smooth," possibly best-in-class, and the model underrated; others are already recommending forks or alternatives like pi.dev.
- Slop accusations — 182 top-level dependencies and 1.3M+ total lines strike some as AI-generated bloat; plus a side-thread on why every lab has converged on TUIs (terminal UIs) instead of graphical ones.