API to Agents publishes its working methods rather than describing them: an open-source repository with the same MCP server implemented in four languages against one sample API, a conformance scenario that CI runs against each of them on every push, a documentation section built from that code, a benchmark of OpenAPI-to-MCP generators run on the same API with results recorded, and pricing rules whose version number appears on every quote. This page collects the evidence in one place.
Four servers, one interface, tested
The mcp-examples repository (MIT) wraps a small shop API in four tools: two reads with pagination, an idempotent write and a guarded cancel. The same four tools, with the same names, parameters, descriptions and annotations, are implemented with the official SDKs for TypeScript, Python, C# and PHP.
A conformance client runs 25 checks against each server, in GitHub Actions, on every push: the tool catalogue, pagination honesty, summarised results, errors that say what to do next, idempotent retries, and a cancel that refuses without confirmation. The conformance page lists every real failure the test caught while the examples were written.
Documentation written from the code
The docs are not a separate artefact: each build guide quotes the repository files it links to, so a reader can diff the page against the source. Concepts, four build guides, connection guides for Claude and ChatGPT, and a production checklist.
Benchmarks we ran, not read
Our generator comparison records what FastMCP and openapi-mcp-generator actually produced from the sample API on 7 October 2026: tool count, names, descriptions, schemas, annotations, error text, and whether the result ran. One of them crashed on its first request. We say so, and we say it may be fixed by the time you read it. Comparisons that only restate feature lists are not published here.
Prices with a version number
Every quote is calculated by deterministic rules, currently version 2026-09-13.v2, from the scope the free audit proposes. The rules are listed line by line on the pricing page: setup from $499, hosting from $29.99 per month, a maximum automatic quote of $3,999. The AI analysis never sets a price.
What is not public, and why
- Customer servers and their tool catalogues: they belong to the customer and carry their brand.
- The audit's analysis prompts: they change as models change, and publishing them would invite gaming the readiness score.
- Internal infrastructure definitions: the security overview describes the design; the Terraform does not need to be public to be auditable by a customer under NDA.
Frequently asked questions
Can I run the examples myself?
Yes. Clone the repository, start the sample API with one Node command, start any of the four servers, and run the conformance test against it. The README has the exact commands; each takes under five minutes.
Is the code I get from API to Agents the same as the examples?
The patterns are the same: credentials server-side, summarised results, idempotent writes, guarded actions, errors that instruct. Your server is designed from your API and your workflows, reviewed by our team, and hosted under your subdomain; the examples are the public, generic version of that work.
Why publish a benchmark that could be out of date?
Because a dated, reproducible measurement is more useful than an undated opinion. The page states the versions and the date, and the method is simple enough to repeat in an afternoon.
Can I see the conformance results?
The GitHub Actions runs on the repository are public. Each shows four jobs, one per language, each running the same scenario.