API Evangelist

MolTrust: "A limit in an agent's system prompt is a request to the model." For Amazon's Ads MCP Server it checks every tool call against a mandate per ad account, and fails closed. https://apis.io/2026/10/26/moltrust-says-a-limit-in-a-system-prompt-is-only-a-request/

Kong runs its review agents in CI, behind its own AI gateway: sandboxed, credentials stripped, approval separate from authorship. And an A/B test found Sonnet runs 40% cheaper on OpenCode. https://apis.io/2026/10/26/kong-runs-its-review-agents-in-ci-behind-its-own-ai-gateway/

What providers say vs. what they do: 987 providers say their writes are safe to retry. Only 428 put an idempotency key in the contract, which is the only part an agent can act on. Measure the difference: https://apis.io/rating/

All 8 Australian banks we profile publish a contract conforming to one government standard. But the CDR reads, it doesn't act, and scams, not open data, are opening up the telcos. https://apievangelist.com/2026/10/26/what-sets-australia-apart-in-banking-and-telco/

The Specification Layer Now Has a Home: one page listing the 52 specifications I read every Monday, from OpenAPI, MCP and A2A to SLSA, CloudEvents and x402, plus the alignment work. Tell me what I'm missing. https://apievangelist.com/2026/10/26/the-specification-layer-now-has-a-home/

A Meter With No Rate On It. Salesforce will charge Flex Credits for registered agent and MCP calls, price TBA. Across 4,986 plans, the market mostly bundles agent access, and where it meters it, it publishes the rate. https://apievangelist.com/2026/10/26/a-meter-with-no-rate-on-it/

Shipping: The Shape Of The Answer. DCSA wrote the ocean shapes as OpenAPI, GS1 wrote the identifiers, and parcel aggregators boiled carrier chaos into nine tracking statuses. The aggregators score ~3x the carriers. https://apievangelist.com/2026/10/26/shipping-the-shape-of-the-answer/

467 providers serve a conformant A2A agent card. Nearly half came from a docs platform default with a single skill: read our docs. Most of the rest are agent-native companies selling company data, trust and on-chain services by the call. Conformant is not the same as useful.

Companies keep asking me whether they need MCP or A2A. It's the wrong first question. Map the flow step by step, own and publish your side, secure it with OAuth and FAPI, put idempotency on every call, and keep as much as you can deterministic.

I profiled 227 of Australia's largest companies ahead of apidays Australia. 117 publish nothing an agent can read. Tech is programmable because it is paid to be; banking and energy because the Consumer Data Right made them. Mining: 41 miners, one Swagger file.

That copy it, beat it, or decline it framing is the part I would keep. Too many new APIs start from a blank spec, when the real work is deciding what is already good enough to copy and what is actually worth beating. Using existing standards as the mold is what keeps that decision honest.

Before you design a new API, iterate an existing one, ship an MCP server or rework your developer presence, profile what your competitors already shipped and turn it into decisions: copy it, beat it, or decline it. Then use existing standards as the mold, so you aren't inventing an interface.

ToolJet MCP built a payments incident room from one prompt: commanders override, engineers add updates, viewers read. The app separates roles; the API it was built through uses one key. https://apis.io/2026/10/13/tooljet-builds-an-incident-room-from-one-prompt-and-its-record-lights-five-dimensions/

GitLab's MCP server is GA: fewer tools doing more, toolsets to trim the catalog so agents pick wrong less, and handoffs to GitLab's agent flows that pause for human approval. Kin 81.0 https://apis.io/2026/10/23/gitlab-says-an-mcp-server-gets-agents-what-they-need-in-fewer-calls/

SerpApi's new eBay Seller API turns any eBay store into JSON: listings, auctions, feedback. Its published contract names one generic search endpoint, so an agent has to read the blog to find eBay. https://apis.io/2026/10/14/serpapi-adds-an-ebay-seller-api-and-its-own-contract-maps-three-operations/

I read all 884 runtime policies across 12 API gateways. AI is now the biggest area, 120 entries, and about two dozen are cost controls protecting the provider's own model bill. Charging the consumer for value: 10 entries at 4 gateways. The cost side got built. The revenue side didn't.

Insurance agreed on its shape fifty years ago: ACORD defines claims, policies, placements and the next-generation digital standards. But the schemas are members-only, behind a login. Of 17 providers tagged ACORD in the catalog, only two score above 40. The shape exists. You just can't easily see it.

We didn't skip hypermedia, we unbundled it. Links became Arazzo, forms became OpenAPI and JSON Schema, meaning became Overlays, discovery became well-known paths and agent cards. The one thing the documents can't carry is what's possible right now, for this resource. That stays in the response.

How programmable is your API? The Kin Score reads six facets and nineteen agent-readiness dimensions, and publishes one number for more than 27,000 API providers. Find yours: https://apis.io/rating/

@apievangelist.com Publishing the enforced profiles would help. The harder boundary is whether a note records medication use, proposes a change, or authorises an order. All can yield valid resources. A repeatable extraction still needs to preserve that distinction before downstream systems act.

API monetization is 100 times more important than the attention it gets, and the word is part of why it falls flat. It sounds like a tax, it only runs one way, and it ignores the cost side, which AI just broke open. Value exchange was the fix. With agents, I think it's API economics.

OpenAI donated Open Responses to the Agentic AI Foundation, a neutral home for a common LLM API. The baseline it walks into: of 120 OpenAI-compatible providers with a spec, 105 declare chat completions and 26 declare responses. The Anthropic Messages dialect is the second camp.

The Train Travel API example in the API Commons moved to OpenAPI 3.2: nav tags, named servers, a QUERY method for station search, a SUBSCRIBE verb through additionalOperations, and dataValue examples. A real document tool builders can test 3.2 support against. Thanks to Polo at Bump.sh.

If your business depends on OpenAPI, please become a member of the OpenAPI Initiative. 32 active members pay for OpenAPI, Arazzo and Overlays, the specifications much of the API economy runs on. 68 companies were members and aren't today, API Evangelist included. Here's the list.

Leonard Richardson's 2008 QCon talk: links tell you where the resources are, and forms tell you what you can do to them. Providers stopped one step short of hypermedia, because only a human at a browser could decide what to do next. Agents are the client that talk was missing.

Atlassian MCP now connects Bitbucket over OAuth: Jira says why, Confluence says how, Bitbucket holds the code. Our catalog missed the server's move to v2. That's on us, filed. https://apis.io/2026/10/22/atlassian-opens-bitbucket-to-ai-clients-over-oauth-and-the-record-has-not-found-the-server/

The Kin Score's third dimension reads each provider in its market, from registers, marketplaces and what providers publish about themselves. Open markets integrate through workflows, regulated markets through standards. Here is where it stands, rules, numbers and open questions.

Nobody in payments wants a creative answer. Turn an invoice into a credit transfer, a remittance into a reconciliation entry, a complaint into a dispute. ISO 20022 and open banking already wrote the shapes down. Plaid scores 84.6; several big US banks barely register for agents.

Kelly Rogers' definition from our Tyk LEAP panel in New York: an agent is ready for production when someone who did not build it can tell you what it does, who authorized it, and how to shut it off. And why banks lead on agents: regulation built the muscle a decade ago.

Foreca promises weather API docs you'll love: uniform JSON everywhere, a live console, a key in two minutes. But predictable data models are a schema claim, and no schema is published. Kin 5.6 https://apis.io/2026/10/22/foreca-says-api-documentation-youll-love-and-publishes-no-contract/

Are you telling me we have to build AND document? Pffft.

Every call is an ungoverned path unless something inspects it, and nothing can inspect a path it hasn't found. My keynote at Tyk LEAP ON TOUR in New York: 6,968 MCP entries, 4,041 real servers, 3,321 reachable, and only 6.7% of providers publish the auth metadata agents need.

The first formal draft of the OpenLint Ruleset Specification is ready for feedback. Spectral rulesets were never specified, only implemented. The draft has 94 testable requirements, conformance tests for each, and 72 open questions that need more than one voice.

A banker told me finance doesn't have enough MCP servers, so I asked the catalog. Of 2,975 finance companies, 245 ship a server that reaches the API, about 8%, and most aren't banks. Of 133 Fortune-tagged institutions, 120 have no MCP at all. They are right about the banks.

A response to Mike Goerlich at Curity on Security Sovereignty: you cannot be sovereign over trust that was never made explicit. Before trust can be portable it has to be visible, as scopes, security schemes, resource metadata and access rules an agent can read and a company can carry.

Job postings don't name AI companies. They name AI products. Across 343,299 postings from 984 companies, 380 name Claude and 142 name Anthropic. 259 say Claude and never say Anthropic. ChatGPT is in 287, OpenAI in 245. A posting asks for the thing you touch.

Five weeks after the MCP Skills Extension went final, I asked every MCP server in the catalog: 9,170 endpoints, 1,450 answered anonymously. Four companies serve skills over MCP the way the spec says: Hugging Face, Glean, Renoolab and BestPrice. We serve 20, and I'm not counting us.

Healthcare has done the most work writing its shapes down: FHIR, US Core, Da Vinci prior auth. The job is turning a fax or a clinical note into valid FHIR resources a server will accept. 169 providers in the catalog say they speak FHIR. Only five have contracts for claims.

Agents usually guess a URL and curl it. Cloudflare's new web search via AI Gateway lets them search instead, only from crawlers that meet its Verified Bots rules and link to the source. https://apis.io/2026/10/20/cloudflare-gives-agents-web-search-through-ai-gateway-with-verified-crawlers-only/

Marketing automation on apis.io: 92.7% document a rate limit, 15.3% idempotency, 3.6% a dry run. It scores 51.7 to 19.6 for the rest of the catalog and cannot promise a retry sends once. https://apis.io/2026/10/21/marketing-automation-tells-agents-how-fast-to-send-and-not-how-to-send-once/

Mike Amundsen moved from asking what a program needs to do next, to what it needs in order to decide. That is context engineering done once, by the API provider: contracts, workflows, schemas, rate limits and access rules written down as explicit facts, relations and capabilities.

Playwright CLI uses fewer tokens per call. Slack measured a full flow: MCP ~40 turns, 3.8M tokens; CLI ~85 turns, 6M. TestMu AI's point: the unit of cost is the task, not the call. https://apis.io/2026/10/20/testmu-ai-says-cli-is-cheaper-per-call-and-slack-measured-it-cost-more-per-task/

A listing says a company takes part. A certification says someone outside checked: Certified Kubernetes, FedRAMP, Peppol, OPC UA. Both belong in a market, not with the same weight. And a bug in our depth code had been quietly erasing every certified vendor's certification.

We re-probed every A2A agent card the catalog tracks: 389 of 487 are live, 62 drift from the spec, 27 are gone. We host 13 cards ourselves for other companies, and held every one out of the market. A provider gets credit when it publishes its own card, on its own domain.

We checked 15,094 SDK packages across 3,802 API providers on npm, PyPI, Go, Maven, NuGet, RubyGems and crates.io. 5,008 are official. Roughly 8,900 that look official are not. Agents install packages, and if the official one isn't clearly yours, they may pick the better README.

NVIDIA tested 65 real DOCA prompts: agents met 19% of the checklist without skills, 100% with them. Its own record shows no agent skills, because nothing points at the repo. Our gap, filed. https://apis.io/2026/10/14/nvidia-measures-agents-at-19-percent-on-doca-without-skills-and-100-with-them/

"We can't have a public portal or public API docs." Your competition already does. Restaurant POS: NCR Voyix scores 27.5, Lightspeed 83.8. The gap is everything after you find the API: pricing, status, sign-up and limits an agent can read without a meeting.

DeepSeek open-sourced DeepInfer: batch-invariant inference, deterministic sampling, and output constrained to a JSON schema, regex or grammar. That moves us past generic decision models to answers shaped for real business outcomes. The shapes already exist in the APIs you depend on.

Naresh Jain had agents build the same service with and without a harness, judged by 178 hidden tests. The failures were at the seams, not the reasoning. What kept a cheaper model honest was OpenAPI and AsyncAPI contracts a machine executes against, not a prose spec.

Cloudflare's new OHTTP Gateway: a relay sees who is asking, the gateway sees what, no one sees both. Its own relay broke that for customers behind Cloudflare, so it built the other half. https://apis.io/2026/10/19/cloudflare-builds-an-ohttp-gateway-so-no-single-party-sees-both-halves/