AI assist
AI assist is optional, off by default, and bring-your-own-key. It is the direction where Tinspec calls a model for you. The other direction — your own agent calling Tinspec — is Agent access.
Turn it on
Section titled “Turn it on”- Settings → AI. Enable it. The switch carries explicit copy about data leaving the machine; read it, because that is the whole decision.
- Add your Anthropic API key. It is stored in the app’s encrypted store and is never handed back to the UI after saving.
- Pick a model.
Until both the consent switch is on and a key is stored, every AI entry point is hidden — not disabled, hidden. There is no AI surface in the app you have not opted into.
What it does
Section titled “What it does”Ask — a question about the API you are working with, answered as streamed markdown. “What does this endpoint actually do?”, “Which of these endpoints create something?”, “What is the difference between these two parameters?”
Draft — describe a request and get a structured one back, built from the spec, opened as an unsaved tab for you to check and send.
Entry points: the AI panel in the right dock, the Ask AI action in the Docs view header, and the command palette (Ask AI About This API…, Draft a Request with AI…).
What leaves your machine
Section titled “What leaves your machine”This is the part worth being precise about.
Only an OpenAPI subset. The context sent with a prompt is a minimal document built from the endpoints of the tab you are on — or from your single spec-backed provider if there is one. Operations, parameters, schemas. That is it.
Never response bodies. Whatever your API returned stays local. The model is not shown your data.
Never secrets. Not environment values, not keychain entries, not tokens.
And you do not have to take that on trust: the AI panel has a collapsible “context sent with prompts” preview showing the exact payload, verbatim, before you send it. Open it once and see for yourself.
Where it runs
Section titled “Where it runs”The LLM layer lives in the desktop app and talks to Anthropic directly with your key. It is deliberately not in the engine — the engine stays headless and remains the single choke point for executing requests, not a chat surface.
There is no Tinspec-hosted inference tier. Nothing routes through Tinspec’s servers, and there is no Tinspec account involved.
Limits
Section titled “Limits”- Anthropic only at v0.1.0-preview.9. The provider boundary exists in the code for other backends, but this release ships one.
- Comprehension and single-request drafting only. There is no agentic loop here — the model does not send requests or read responses. If you want an agent that drives Tinspec, that is Agent access, and it runs on your agent’s tokens rather than a key you give us.
Turning it off
Section titled “Turning it off”Flip the Settings → AI switch and every AI surface disappears again. Delete the key to remove it from the encrypted store.