Grok Build Goes Open Source After Privacy Backlash, But the Repo Won't Accept Your Code
An AI coding tool caught silently uploading entire repositories now ships as open source, but the repo is a read-only showcase with no community contribution path. Developers who audit the source can verify upload behavior themselves, yet the absence of reproducible builds means the binary you download may still differ from what you read on GitHub.
A traffic analysis of Grok Build 0.2.93 caught the tool uploading full Git repos—including unread files, history, and .env strings—far beyond what a coding task required. The `/privacy` command arrived as a quick fix, but researchers found whole-repo uploads persisted even with model-training toggles off. On July 15, SpaceXAI open-sourced the Rust CLI, agent runtime, and tool framework under Apache 2.0 and reset all user quotas.
The repo is genuinely inspectable: you can build from source, examine file reads, sandboxing, and upload logic, and point the harness at custom models or local inference via `base_url`. But the project explicitly refuses external pull requests and unsolicited patches, carries a single commit, and has no tagged release or reproducible-build proof that the binary matches the source. Codex CLI and Gemini CLI both accept community contributions; Grok Build does not.
Cursor doubled Grok 4.5 usage quotas in the same window, turning a privacy crisis into a competitive giveaway. The open-sourcing moves trust from a corporate promise to auditable code, but it does not retroactively fix the earlier uploads or independently verify that the current binary has stopped them.
Open-sourcing the CLI turns a privacy scandal into a trust-repair narrative, but the timing and bundled quota resets make it read more like crisis PR than a philosophical commitment to openness.
Refusing external contributions while applying an Apache 2.0 license creates a strange hybrid: the code is legally forkable, but the upstream project is functionally closed to community development.
A single-commit history and no tagged release mean the repository offers auditability without verifiability—you can read the source, but you cannot confirm the binary you run was built from it.
The competitive response from Cursor, doubling Grok quotas, shows how quickly AI coding-tool vendors weaponize each other's missteps into user-acquisition plays.
I've never understood this narrative. Caught stealing code, then open-source the client. Isn't the most important thing what they actually did with the code, and how to guarantee that no residual copies remain and are completely deleted?