@fberrez1 In truth, that is the best critique to make — OpenAPI-only tooling falls aside on the second one endeavor buyer.
How MCP Bridge layers it:
1. OpenAPI / WSDL / GraphQL as a place to begin, now not the end line. The Device Builder imports what you may have, nevertheless it generates an editable device definition, now not a sealed wrapper.
2. Adapters for reaction normalization — when an API returns 200 with `{“error”: “…”}` within the frame, or other error shapes in line with endpoint, you write a small change into (JavaScript, runs in our sandbox) that maps the mess to a constant contract. The agent sees a blank device; the bridge does the grimy paintings.
3. Multi-step flows as a unmarried device — auth handshake → token cache → primary name → reaction shaping is one composable device from the agent’s viewpoint. The LLM by no means has to reason why about float state.
4. Customized auth handlers — now not simply OAuth and API keys. We now have shipped gear towards APIs that need a nonce signed with a personal key, exchanged for a consultation token, then the actual request. 3 calls and a few crypto — one device to the agent.
5. Failure shaping — inconsistent standing codes get categorised on the adapter layer right into a retry/no-retry resolution, so the agent does not have to determine whether or not a 200 with an error frame way “check out once more” or “surrender.”
Maximum “attach any API” gear suppose the sector seems like Stripe. Undertaking APIs do not. The Device Builder and the sandbox exist in particular to take care of the space between what is documented and what in fact works.
