AI in API tooling usually comes with a quiet trade: your requests, your schemas and your company's API designs get uploaded to the vendor's cloud, parsed on their servers, and sometimes used to train the next model. For early-stage products and internal APIs alike, that is a non-starter.

Powerduck's approach is different: AI works across the entire API lifecycle, while the local file stays the source of truth.

Where AI actually helps in an API lifecycle

Not a chat bolted onto the sidebar — AI that reads and writes the same document you do:

  • Design: describe a resource in plain language and get a proposed operation with schemas, status codes and examples.
  • Import and triage: paste a cURL command or a Postman collection and let AI normalize it into clean, consistent OpenAPI.
  • Test generation: turn an operation into scenarios, assertions and edge cases, including the unhappy paths people forget.
  • Debugging: hand a failed response or a confusing error to the assistant and get a grounded explanation that references your spec, not generic advice.
  • Refinement: tell a proposal "use ISO 4217 currency codes" or "split billing address into its own schema" and watch the plan update before anything is applied.

Plan, review, apply, revert

AI never edits your file by stealth. Every proposal appears as a plan with the affected operations marked. Applying a proposal creates a revision with a full diff, and any revision can be reverted. Your manual edits are preserved: proposals are patches against the current document, not regenerations that overwrite your work.

That review-and-revert loop is what makes AI safe to use on a spec you care about. You stay the author.

Your model, your terms

Model choice is yours:

  • Bring your own provider — configure any OpenAI-compatible endpoint, including self-hosted models, and your traffic goes directly there.
  • Powerduck model packs — a zero-config option for teams that want AI without managing API keys, metered as resource packs from the console.

When Powerduck's hosted model is selected, requests go through a thin relay with a fixed product system prompt; every other provider is called directly from your machine.

No account on your own machine

The desktop application is local-first by design. There is no login wall to open your own files, no telemetry-gated workbench, and no requirement to put your specs in someone else's repository. The browser demo asks you to sign in only because shared cloud resources need per-account rate limits — the downloadable client needs nothing of the sort: configure your own model (or none) and work offline.

Try it on real input

Grab the messiest cURL command from your team's chat, drop it into a fresh spec, and ask the assistant to turn it into a documented, tested operation. Review the plan, apply it, revert it, apply it again — and notice that at no point did the file leave your machine.

One local OpenAPI spec, an AI-native workspace around it. Start with your own API.