Give an agent real APIs

One bridge command turns this whole catalogue into tools for any MCP client. The agent gets the declared schema for each operation, so it calls them with the right arguments instead of guessing at them.

The same meter, the same boundary

An agent calling through MCP is charged exactly as an HTTP caller is, against the same account, with the same capability checks. There is no cheaper back door and no unowned lane — a call that cannot be attributed to an owner cannot be charged to anyone and cannot be refused to anyone.

Toolbanks keep the list short

Handing an agent a thousand tools makes it worse, not better. A toolbank is a named subset — the eight operations this particular agent should have — so the model chooses from a list it can actually reason about while everything else stays one permission away.

Questions

Can a coding agent use this?

That is what the MCP server is for. Point any MCP client at the bridge one-liner and every operation you are entitled to becomes a tool it can call, with the same names, the same schemas and the same meter as the HTTP API.

Is every call metered?

Yes, and deliberately so: every path — HTTP, the MCP server, a chatbot function call, a workflow step — goes through one pipeline that charges before it executes. This engine once ran entirely unmetered, which is free compute with a price list on it, and a test now asserts there is no route around the meter.