xAI Grok Build CLI Uploads Entire Repos and Git History

xAI Grok Build CLI sends complete repository contents and git records to cloud storage after login.

Analysis of network traffic revealed three data streams. The tool reads local files including environment keys. It transmits them in plain text. It also packages every file from connected repositories regardless of prompts.

One test with a 12 GB repository moved 5.10 GiB through storage endpoints. Conversation channels carried only 192 KB. The ratio reached roughly 27,800 times more volume on the storage path.

The mechanism runs by default. Disabling model improvement settings leaves trace uploads active. Session data reaches both xAI endpoints and a Google Cloud Storage bucket named grok-code-session-traces.

Developers reported seeing uploads continue even after explicit instructions to avoid reading files. The behavior matches the injected summary from traffic logs posted on GitHub.

Traffic patterns match session state uploads

POST requests to /v1/responses carry file contents in readable form. Separate calls to /v1/storage deliver session_state archives. Both routes return HTTP 200 responses confirming acceptance.

The uploads occur independently of any AI agent read operation. Entire git bundles move even when the user prompt states "do not read any file". This separation appears hard-coded in version 0.2.93.

No public documentation lists the storage bucket or its retention policy. The analysis did not confirm training use of the data but verified transmission, receipt, and storage.

Default settings keep uploads enabled

Settings endpoint returns trace_upload_enabled set to true after users toggle model improvement off. The flag persists across restarts. No UI switch currently disables the flow.

This configuration differs from many peer developer tools. Those tools tie telemetry to explicit consent toggles. The persistent flag raises questions about user control.

Volume contrast highlights storage priority

Conversation traffic remained under 200 KB in the 12 GB test. Storage traffic reached 5.10 GiB. The imbalance shows the primary payload travels outside normal model interactions.

Developers can reproduce the pattern with standard packet inspection tools. The gist shared on GitHub lists exact endpoints and headers. Independent verification requires only a fresh login and repository clone.

Privacy implications for repository contents

Full git history includes commit messages, author metadata, and deleted file versions. Environment files may contain keys or tokens. Both categories now travel to xAI infrastructure without granular selection.

The company has not published a data use statement covering these bundles. Storage occurs in Google Cloud buckets dedicated to session traces. Retention periods remain unspecified.

Competitive context shows divergent practices

Other CLI coding assistants offer per-project data controls. Some require explicit repository inclusion lists. xAI Grok Build CLI differs by treating the entire working directory as default input.

No comparable tool from major labs has published similar bundle upload behavior in public traffic reviews. The gap leaves xAI as the clearest documented case.

What remains unverified

The posted analysis did not include server-side logs proving downstream processing. It also stopped short of claiming model training use. Those gaps leave open questions about final data handling.

xAI has issued no official response to the gist. Independent audits would require either source code access or additional endpoint testing.

Signals to monitor in coming months

Watch for updates to the settings endpoint flag. A change to trace_upload_enabled would indicate policy adjustment. Monitor release notes for version 0.2.94 or later.

Track community reports on repository size limits or selective upload options. New complaints or confirmations on forums would signal whether the flow persists.

Observe any data processing addendum published on xAI developer documentation. Formal language on retention or training exclusion would close the current verification gap.