Use AI to ask questions about selected Flightops evidence and investigate a problem with a model. Normal attach, launch, and results work without AI.
Start here
- Connect a provider: choose a model and set up credentials.
- Use the assistant: select evidence, ask a question, and review a response.
- Manage credentials and troubleshoot: check a connection, change providers, or disconnect.
- Prioritize evidence with Jev: optionally select a useful receipt before asking your assistant.
- Compare evidence selectors: evaluate the same incidents without assuming Jev is better.
What is available
| Connection | Current scope |
|---|---|
| OpenAI API or Anthropic API | Guided setup, explicit connection test, and assistant requests. Requires a model your account can use. |
| Local or custom OpenAI-compatible endpoint | Advanced setup with your endpoint and model. Compatibility depends on the endpoint. |
| ChatGPT account (preview) | Isolated text assistant for reviewed questions and attached evidence. Requires the pinned Codex runtime, Linux bubblewrap, browser sign-in, and a successful explicit connection test. Native Codex tools are disabled. |
A saved connection does not mean it has been tested. Live ChatGPT account exchanges are not accepted for the beta; see known limits. Use the testing checklist on your own account and machine.
How it works
Flightops starts in Standard mode. Setup saves a connection; flightops --ai or the assistant panel's m key enables AI. Enabling it does not send a prompt.
You choose evidence and write a question. Before sending, the assistant shows the destination and material to disclose. Approve that request with y, or reject it with n. Further retrievals and actions have their own review.
An assistant response is an explanation or proposal, not a passing test result. Use the retained source evidence and an explicit check to verify a claim. Hardware actions remain unavailable.