7,411 third-party posts across 1,608 provider blogs between September 27th and October 3rd. Cloudflare opened a gateway that lets you charge agents for API and MCP calls, AWS shipped an agent skill for metered billing, and Pipedrive priced every one of its operations in its OpenAPI. On the other side of the ledger, WorkOS counted three MCP authentication bugs in thirty days with one root cause, and a survey found most companies that enforce agent permissions still let their agents share credentials. Arcade rescored 90,000 MCP servers and found 0.28% earn an A. And I re-scored twenty-six thousand companies before flying to London.

6,061 third-party posts across 1,548 provider blogs between September 20th and 26th. The week I called the gateway a tired concept, Cloudflare took MCP server portals GA, Kong shipped an MCP server for every developer portal, DigitalOcean pitched an Action Gateway that holds the agent's credentials for it, and Google turned its API Gateway into an MCP server driven by an OpenAPI extension. Meanwhile WorkOS made the sharpest point of the week — delete the MCP servers and the agent is back to a long-lived key in a shell — and Arize replaced fifty-eight tools with two.

7,556 third-party posts across 1,558 provider blogs between September 13th and 19th. This week the agent conversation stopped being about capability and started being about the invoice — Redocly measured an 80% cost reduction from letting the model write code instead of chaining tool calls, and Vonage found a token tax hiding inside your JSON Schemas. Meanwhile Meta shipped MCP for WhatsApp Business and the agents still do not get their own identity. And the MCP post count went up only because one company published twenty-three help pages in two days.

7,424 unique posts across 1,787 provider blogs between September 6th and 13th. Last week four vendors shipped agents an identity. This week researchers checked what those identities are actually allowed to do and found one in five MCP access policies broken or missing — while agents claiming to be OpenAI systems wrote seventeen thousand posts to a wiki that was not supposed to accept writes. Meanwhile a second way of connecting an agent to your API showed up, and it is not a protocol.

8,601 unique posts across 1,971 provider blogs between August 30th and September 6th. Four identity vendors shipped agent-identity product inside a seventy-two hour window, and in the same seven days at least four other companies published the argument that identity is the wrong layer to solve this at. MCP finished its migration from the engineering blog to the support desk and started a second one, to the security desk. And OpenAPI was named in exactly one third-party post title all week.

6,768 unique posts across 1,927 provider blogs between August 23rd and 30th. Last week consent grew a scope selector. This week the enterprise version of it went generally available, and the harder problem showed up right behind it — a grant is one hop, and agents make several. Meanwhile OpenAPI reappeared in third-party headlines for the first time in a fortnight, and it reappeared because somebody trojanized its codegen. And one in five MCP posts this week was not an announcement at all. It was a support article.

6,583 posts across 1,735 provider blogs between August 15th and 22nd. For three weeks the industry has been working out how an agent proves who it is. This week it moved on to the harder half of the question — how much that agent is allowed to do — and shipped it as product: task-based OAuth consent at Cloudflare, custom scopes at Clerk, enterprise-managed cross-app access at Auth0, step-up challenges at WorkOS, identity propagation from AWS through to JFrog. Meanwhile 76 posts named MCP in their title and not one third-party post in the whole network named OpenAPI, AsyncAPI, GraphQL, JSON Schema, Arazzo or gRPC.

5,806 posts across the 5,299 API-bearing providers I track between August 7th and 15th. Last week the identity vendors reached into a drawer of old RFCs to answer who is calling. This week the network layer answered a different question — what is this traffic? — and answered it by inspecting headers and guessing. Cloudflare Gateway now detects MCP by protocol heuristics, Fingerprint will tell you which AI tool called your server, and OWASP has an MCP Top 10. Meanwhile Google did the opposite thing and put AI model routing inside an OpenAPI document, where it is declared rather than inferred. That contrast is the whole issue.