Data Usage
What CrabCode collects, what it doesn't, how to turn telemetry off, and where local data lives.
What it is
CrabCode is the client that orchestrates model calls, your prompts, and tool execution. This page documents what touches your disk and what crosses the network.
Model requests (the core data flow)
When you send a prompt, CrabCode forwards the conversation, project context you've authorised CrabCode to read, and tool results to the model provider:
- By default through the acosmi gateway at acosmi.com / acosmi.ai, which routes to the underlying provider
- You can also configure a direct third-party provider connection (concrete names come from the gateway's ListModels response)
Retention and training policies for model requests are set by the provider. See the acosmi.com privacy page or the third-party provider's docs for the authoritative policy — we don't duplicate them here to avoid drift.
Telemetry and quality surveys (opt-out)
By default CrabCode also sends:
- Anonymous telemetry: command usage frequency, performance metrics, error counts (product improvement)
- Error reports: model API errors and tool failure stacks (common secrets are stripped)
- Quality surveys: occasional one-line "did this help?" prompts at session end
/bugreports: uploaded only when you run the command yourself
Turning telemetry off
| You want to disable | Do this |
|---|---|
| All telemetry (keep update checks etc.) | DISABLE_TELEMETRY=1 |
| All nonessential network traffic (includes auto-updates, model capability refresh, release notes, …) | CRABCODE_DISABLE_NONESSENTIAL_TRAFFIC=1 |
/bug command | DISABLE_BUG_COMMAND=1 |
At the essential-traffic level CrabCode keeps only the network calls required for model requests and sign-in.
Local data
CrabCode stores under ~/.crabcode/:
- Global config (
~/.crabcode.json) - Per-project state (one subdir per project, slug from cwd)
- Session PID files (concurrent-session detection)
- Team config / tasks / members (if you use teams)
- Debug logs
These stay only on your machine and never upload — unless you run something like /bug yourself.
ZDR (Zero Data Retention) organisations
Organisations on a Zero Data Retention plan:
- Model requests are not retained long-term server-side
/bugreports are refused- Some surveys are auto-disabled
Exact scope is set by your org admin and the acosmi.com privacy page.
Limits and caveats
- Direct third-party providers: when wired to Bedrock / Vertex / Foundry directly, your prompts go to that cloud — CrabCode and acosmi.com don't see them
- Third-party MCP servers: data sent through an MCP server is governed by that MCP, not acosmi.com
- Crash reports include stacks but not source content: file paths appear; file contents do not
/buguploads the transcript — you'll be asked to confirm; it contains your current session