Solo.io’s Christian Posta has written a practical case against mapping an OpenAPI one-to-one onto MCP tools, and it starts from a line worth keeping: “We’ve invested 20+ years in building useful, reusable, valuable APIs, should we really be reinventing everything just because a fancy new protocol showed up?” His answer is no, but he is equally blunt that APIs “were written for ‘humans’” who already know the task, and that chaining calls through a model is brittle. Pushing “a 10,000-row response through the context window just to feed the next call wastes tokens and increases error rates.” agentgateway offers three ways out: direct exposure, curated tools with CEL-shaped responses and chaining, and a “code mode” that exposes a single run_code tool. The agent writes a short JavaScript program, the gateway runs it in a sandbox against typed functions generated from the OpenAPI, and “intermediate API responses never cross the model’s context window.”
There are no vendor performance figures in the post, which makes it easy to assess. The argument holds up without them. A Backend-for-Frontend layer built for agents is a sound pattern, and the details are the useful part: flatten path, query, header and body parameters, keep credentials away from the model, and shape the response before it comes back. What the post does not discuss is the cost of code mode. The agent now writes executable code against your APIs, so the description it gets is the only guardrail on what that code may do. A typed function signature says what a call takes. It does not say whether the call is safe to retry or whether it can be undone.
The catalog shows the gap. The Solo.io provider page lists 15 API pages, and they describe Solo’s developer-portal surface rather than the gateway in the story: the Solo.io API Products API, Solo.io API Keys API, Solo.io OAUTH Credentials API and Solo.io Subscriptions API. Those four are what you would manage when you put curated APIs in front of consumers. The run_code tool, its sandbox and the CEL transforms don’t appear as a contract anywhere in the record.
Solo.io has a Kin Score of 60.4, in the strong band. Developer ergonomics at 78.6 and discoverability at 76.7 carry it, and contract governance at 28.0 holds it back. Its Agent Readiness is 34.8. Nine dimensions are lit, including the MCP server, error semantics and OpenAPI examples. Twelve are unlit, among them idempotency, reversibility, dry-run mode and pagination. In code mode those are exactly the properties an agent needs to read before it writes a loop. Posta is right that the gateway should do the orchestration. His own estate shows the next step: the APIs being orchestrated have to say what is safe to run.