Behavior-verified API integration
Paste an OpenAPI spec, a Postman collection, or one cURL command. DocentAPI turns it into typed tools, works out which call produces the id the next one needs, runs the read-safe operations against your live service — and writes down what actually came back, with the evidence attached.
No signup · keys never stored · 24-hour anonymous workspace
02Surfaces
Humans read docs. Agents need tools. Most APIs can answer for only one of them — and the agent traffic is already arriving. DocentAPI answers for both, from a single import.
For humans
A hosted, shareable page for your API — real docs, working examples, and a playground where visitors test calls with their own key. Claimable by you, linkable by everyone.
For agents
The same import, exposed over the Model Context Protocol. Claude, Cursor, and Copilot call it directly — typed tools, auth handled, every call checked against how your API really behaves.
03Lineage
An agent that invents an identifier fails, retries, and fails again. DocentAPI reads every operation’s output against every other operation’s input and works out what feeds what — weighing eleven signals, from a shared schema title down to a type mismatch that argues against the link.
≥ 0.95
Measured against four hand-labelled corpora built to the structural shapes that break this — RPC paths with no resource hierarchy, pagination params, five resources all exposing a bare id. The build fails below the gate.
Silence
Recall is measured and printed, never asserted. Asserting on it would pressure the engine toward guessing, which is the one failure this is built to avoid. A low-confidence edge is withheld, not shown with a hedge.
Every edge
No link is asserted without the signals that produced it. A pet’s id and a category’s id are both bare integers named id; resolving which is which is most of the work, and the reasoning is published with the answer.
04The truth layer
The single most expensive question when integrating an API is “what do I put here?” — and a spec answers it for almost nothing. DocentAPI classifies every writable field into one of 5 origins, and records how it knows.
| Origin | What it means for a caller | Decided by |
|---|---|---|
server_generated | The API produces it. Anything you send is ignored or rejected. | readOnly in the schema |
constant | Exactly one legal value, so there is nothing to decide. | const in the schema |
produced_by_api | Another operation returns it — and we name which one, per field. | Lineage |
enum_constrained | Pick from a fixed list. We can show you the list. | enum in the schema |
caller_supplied | Genuinely yours to choose. This is the only one that needs you. | Nothing else matched |
Each annotation also carries how confident we are entitled to be. A value the owner confirmed is marked human; one inferred from a quote in their own documentation is assumed, and carries that quote; one we worked out from structure alone is heuristic. Only a person can set the human mark — a database constraint makes anything else unrepresentable, not merely discouraged.
Whether the server honours the id you send. Whether a PUT replaces the record or merges into it. What userStatus: 2 means. Nobody can derive these — so we ask you, once, and we never ask with a blank box.
Every question is one of 8 shapes, and each shape’s answers are enumerated from the field itself — its type, its enum, its position in the path, what lineage already found. The options are not generated. They are what the structure allows.
Ordered easiest first, because someone who answers three quickly keeps going. One answer applies to every operation the field appears on, and skipping is a real answer — it publishes an honest unresolved rather than a guess.
Who assigns this id?
add_pet · body.idadd_pet creates a record, and "body.id" is writable, so a caller cannot tell whether sending an id does anything.
Unlocks Agents stop inventing ids on create, or stop omitting one you actually require.
05Verification
Transpilers echo the spec and hope. DocentAPI executes the tools, catches where the docs lie, patches the tool definitions — and re-verifies so drift never reaches an agent.
Session replay — scripted from a real drift finding.
Stainless generates SDKs. Mintlify renders docs. Speakeasy transpiles MCP. DocentAPI verifies the surface agents actually touch — and scores it.
Worked example — not a measurement of any API.
06Score
A 0–100 grade of how well an agent can drive your API. Two of the four sub-scores are earned by making real calls against the running service; two are graded from the document. We label which is which, on the page and in the MCP tool that explains it.
Live — a read-safe call is deliberately malformed
Do failures explain themselves, or dead-end at an unlabelled 400?
Live — real responses compared against the documented shape
Do the top-level keys and types match what the spec promised?
Static — graded from the declared scheme
Can an agent discover and satisfy auth without a human in the loop?
Static — whether a retry key is offered at all
Agents retry. Does the API give them a safe way to?
Bands 90+ excellent · 75+ good · 55+ mixed · 35+ weak · below that, the spec alone is not enough to integrate reliably.
07Deliverables
Paste a spec once. DocentAPI fans it out into 6 surfaces — generated together, verified together, and kept in sync with every change you ship.
A public, shareable page for your API — docs, examples, and a try-it playground. Every page is a landing page for your API, findable and linkable.
Visitors test real calls with their own key — used for that one request, never stored, never logged. No signup, no sandbox to provision.
www.docentapi.xyz/mcp/you — a drop-in endpoint Claude, Cursor, and Copilot call directly. Typed tools, auth handled, zero infrastructure on your side.
A 0–100 grade of how well agents can drive your API — measured by executing the tools, not reading the spec. Your number to beat.
“Agent-Ready 87” in your README and docs — proof your API works for agents, and a live link back to your page for everyone who sees it.
Prove you own the domain — DNS, meta tag, or email — and the page is yours. Run read-safe probes and the score turns from preview to verified.
Every public page, badge, and claim link is a door back to your API. Adoption compounds while you sleep.
Every public API gets a live page — even before its owner shows up. Found yours? Prove you own the domain and make it official.
Drop it in your README and docs. It renders your live score — and links every reader back to your integration page.
Visitors try your API with their own key. It’s used for the one call and discarded — never stored, never logged, zero sales calls.
Your users paste one URL into Claude, Cursor, or Copilot and their agents are calling your API — metered, rate-limited, every tool annotated read or write.
Pricing
Public pages are our distribution and your adoption. Plans cover private APIs, teams, and business-critical reliability.
The importer at the top is live — no signup, no credit card. Leave your email and we’ll send your claim link plus each verification feature as it ships.
Free for public APIs · anonymous workspaces expire in 24 hours unless claimed