Riviera documentation
Riviera turns an API you already have into tools AI agents can use, serves them from an endpoint you don't build or host, and records every call. These pages say what it supports today, including what it doesn't.
Concepts
- Project
- One agent-facing interface: its tools, versions, access settings and call history.
- Tool
- One thing an agent can do. A tool has a name, a description, an input schema and, when your API declares one, an output schema. Agents choose a tool by reading its name and description.
- Draft and Production
- You edit tools in Draft. Production is the version agents read. Nothing in Draft reaches an agent until you publish.
- Version
- Publishing creates a new version and makes it Production. Every published version is kept exactly as it was, so you can see what agents were reading at any time.
- Endpoint
- The address agents connect to. Riviera serves it. When an agent calls a tool, Riviera calls your API, using the credentials you saved, and returns the result.
- Tool call
- One recorded call: the tool, the version that served it, the outcome, how long it took and, when your API answered with an error, its HTTP status. Request and response contents are not captured by default.
Two ways to start
Import an API. Give Riviera an OpenAPI document and it serves the operations it can map as tools. There is no MCP server for you to build or host. This is the path these docs describe in detail: importing an API.
Connect an MCP server you already run. Riviera discovers its tools, lets you edit what agents read, and passes calls through to your server. Only tools are managed. Prompts and resources are not. This path has no page of its own in these docs yet.
Both end in the same place: a project with editable tools, versions, an endpoint and recorded calls.
In these docs
- Importing an API
Which OpenAPI documents Riviera imports, how an operation becomes a tool, every reason an operation is excluded, and how to edit and publish the tools.
- Authentication
How Riviera authenticates to your API and how agents authenticate to Riviera: API keys, bearer tokens, OAuth client credentials and per-user OAuth, with their limits.
- Connecting an AI client
Connect an MCP client to a published Riviera interface: the endpoint, access tokens, per-user sign-in, the clients we have tested and the limits of that testing.
For a technical look at what real OpenAPI documents contain, read what converts cleanly, and what doesn’t.
Testing and calls
Model tests let you write the requests your customers make and the tool that should handle each. Riviera asks a real model which tool it would choose, on Production and on your draft. A test checks tool selection only: it does not call your API and does not grade the arguments the model would send. Each plan includes monthly runs: 200 on Free and 2,000 on Pro. Tests can be temporarily unavailable when Riviera’s own monthly spending safeguard is reached.
Activity lists every tool call with its tool, version, outcome and duration. When your API returns an error, the agent is told the call failed together with its HTTP status. The error body is not passed on.
Limits
Free includes 1 project, 10,000 tool calls a month and 7 days of call history. Pro includes 5 projects, 100,000 tool calls a month and 30 days of call history. At the monthly limit, calls stop before they reach your API until the next UTC calendar month or a plan change. See pricing.
Questions these pages don’t answer? Email hello@riviera.so.