Grok Build CLI
A terminal-based AI coding agent CLI tool released by xAI, used for AI-assisted code development and conversation within a command-line environment. It employs a dual-channel network architecture separating model conversation from data storage.
Timeline (last 90 days)
xAI's Grok Build CLI 0.2.93, after receiving only an "OK" reply and without reading any files, still pushed a complete git bundle containing 4 commits and 47 files to xAI's storage endpoint.
In a test with a repository of about 12GB of random files, about 192KB flowed to the model channel while about 5.10GiB flowed to the storage channel, split into 73 chunks of about 75MB each.
Grok Build CLI's model conversations go through claude-chat-proxy.grok.com/v1/responses, storage goes through /v1/storage on the same host, and there are multipart initialization and batch_exists checks.
After turning off the "Improve the model" privacy switch, the git bundle upload still succeeded with HTTP 200, and the server's /v1/settings response still carried trace_upload_enabled: true; this switch controls training, not uploading.
In a real task where the agent actually read files, the contents of a git-tracked .env file entered the model conversation turns, and the same batch of values also landed in the session_state archive being prepared for sending to storage.
On July 13, 2025, the same Grok Build CLI version 0.2.93 suddenly stopped storage requests; six retests all showed zero /v1/storage uploads, and the server returned disable_codebase_upload: true and trace_upload_enabled: false.
After testing the /privacy function, cereblab concluded: it controls data retention, not what data is sent.
A deny rule like Read(secret.txt) can prevent the agent from reading the file, but the file tracked by git still exists in the uploaded git bundle.
All Facts (8)
Grok Build CLI's model conversations go through claude-chat-proxy.grok.com/v1/responses, storage goes through /v1/storage on the same host, and there are multipart initialization and batch_exists checks.
50%UnverifiedAfter testing the /privacy function, cereblab concluded: it controls data retention, not what data is sent.
50%UnverifiedA deny rule like Read(secret.txt) can prevent the agent from reading the file, but the file tracked by git still exists in the uploaded git bundle.
50%UnverifiedAfter turning off the "Improve the model" privacy switch, the git bundle upload still succeeded with HTTP 200, and the server's /v1/settings response still carried trace_upload_enabled: true; this switch controls training, not uploading.
50%UnverifiedIn a test with a repository of about 12GB of random files, about 192KB flowed to the model channel while about 5.10GiB flowed to the storage channel, split into 73 chunks of about 75MB each.
50%UnverifiedIn a real task where the agent actually read files, the contents of a git-tracked .env file entered the model conversation turns, and the same batch of values also landed in the session_state archive being prepared for sending to storage.
50%UnverifiedxAI's Grok Build CLI 0.2.93, after receiving only an "OK" reply and without reading any files, still pushed a complete git bundle containing 4 commits and 47 files to xAI's storage endpoint.
50%UnverifiedOn July 13, 2025, the same Grok Build CLI version 0.2.93 suddenly stopped storage requests; six retests all showed zero /v1/storage uploads, and the server returned disable_codebase_upload: true and trace_upload_enabled: false.
50%