Your agent read the docs. It still called the API wrong.

A human-shaped spec becomes a tool your agent picks by intent.

Point Gecko at any OpenAPI — no provider help. It returns a handful of first-call-correct tools the agent chooses by intent, not dozens of raw operations it has to reverse-engineer. When the spec misdeclares its auth, Gecko resolves it here, before a call goes out.

gecko.search("find the latest match odds") → prepare → call(mode="recorded")

• Unilateral ingest — point at any spec and get intent-shaped tools; zero provider cooperation, no integration code.
• Auth misdeclared in the spec (a static header typed as oauth2) is caught at comprehension time, not on a failed live call.
• Auth is hidden from the agent — credentials injected at call time, never exposed in a tool def.
• Operations the token can't reach are filtered out, so the agent can't pick a call it can't make.

Prove the call before you spend a cent.

recorded mode synthesizes a response straight from the schema. $0, fully offline. Test that the call is correct before any live request touches the API.

• Deterministic examples generated from the schema — no network, no key, no spend.
• One code path for recorded and live; they diverge only at the transport edge.
• Falsify the integration offline first — live smoke is the final check, never the debugger.
• $0 before you commit to a paywalled API — evaluate the fit before you subscribe.

Tests that prove first-call-correctness.

We're building gecko-testgento auto-generate TDD tests that assert the agent's call matches the spec — and regenerate them when the API drifts.

• Generated tests assert the prepared call matches the surface — green means first-call-correct.
• Re-comprehend on drift: a renamed field regenerates tools and tests, not a 2am hotfix.[Coming · V2]
• Tests are a pure function of the surface — versionable, reviewable, diffable.

A hosted MCP your agent adds in one line.

Add Gecko to Claude or Cursor with a single command. Gecko handles the access handshake and injects credentials at call time — invisible to the agent — then the agent calls the real API directly. We are the control plane, never the data plane.

• Streamable-HTTP MCP endpoint — one-click add to Claude or Cursor.
• Gecko handles the access handshake; credentials injected at call time, never custodied or logged. Not a payment rail — your agent calls the real API directly.
• Control-plane only: we store the API surface and generated tools, never a response payload, user data, or a secret. That's the promise that lets us ingest any API unilaterally.
• Drift-resilience we're building: re-comprehend instead of re-integrate when the API changes.[Coming · V2]