AI policy

Codexa has a provider abstraction for Anthropic and OpenAI models. The operator of your instance decides which keys are configured and which model is the default. If no key is configured, AI features are disabled and the interface says so; nothing is sent anywhere.

• Your prompt and the conversation history for the current project conversation.
• The project files the agent chooses to read, and the listing of file paths.
• Build, test, and type-check output when the agent validates its changes.
• Not sent: your password, session, environment variable values, integration secrets, or files from other projects.

Codexa does not train models on your data. Whether a provider retains or uses data sent through its API is governed by that provider's terms for API usage and by the operator's agreement with them.

• A snapshot is taken before every run; you can review per-file diffs and revert.
• Runs stream their plan and tool calls; you can cancel at any time.
• Repairs are bounded to two attempts so the agent cannot loop indefinitely.
• Usage (requests and tokens) is recorded per workspace and enforced against your plan; see credits.

Generated code can be wrong, insecure, or infringe on others' work. Review it as you would any contribution before deploying it to users. Follow the configured provider's usage policies and the acceptable use policy.