Connect your own provider
The agent uses the account you choose: Claude through subscription OAuth, Codex and Grok through device code, or an OpenAI- or Anthropic-compatible endpoint through a key and base URL.
Keys and tokens stay in Keychain. There is no Odete account or server mediating the request; it goes directly to the provider you selected.
Choose a mode before asking
Chat reads the project and answers without editing. Plan writes only .odete/plan.md, which is useful for organizing work. Build can edit the project, always through patches you can review.
The agent can read files, list, search, inspect the terminal, run shell commands, replace or write content, and talk to the GitHub API, according to the chosen mode and permission.
- Ask: request approval for sensitive actions.
- Auto: automate within the permission you granted.
- Full: still asks before writing to someone else’s repository.
Keep human review
Every edit arrives as a patch. Accept, reject, or accept hunk by hunk, then compare the result with tests and Preview. A checkpoint is written before each turn so you can undo the last round.
Ask for an investigation or a plan first when a change crosses many project areas. The agent does not replace security review, integration testing, or product decisions.
Understand the context sent
When you send a message, your prompt, files the agent requests, and needed command output may be sent to the AI provider. Your code and conversation are governed by that provider’s privacy policy.
Conversations, patches, checkpoints, and plans stay in .odete/ in the project and are excluded from Git. Do not send secrets the provider should not receive; reduce context and review permissions before using Build.