MCP for AI agents

EvidentSource has a Model Context Protocol server. An AI agent connected to it can:

• discover the databases on the server, and read guides written for agents;
• read State Views, events and the transaction log;
• run commands.

Publishing a component is an HTTP multipart upload, not an MCP call: a compiled component is too large to pass through a tool call. The evidentsource://guides/publishing resource explains how.

Endpoint

https://mcp.sandbox.evidentsource.com

No authentication is required on the public Sandbox. (See About the Sandbox.)

In an agent config

Add the endpoint to your agent’s MCP configuration as a remote, HTTP server. For the exact format, see your agent’s documentation, or modelcontextprotocol.io.

What the agent sees

• Two tools. fetch makes GET and POST requests to the HTTP API. execute_state_change runs a JSON command through the /latest State Change route. Commands can also use fetch with the State Change route.
• Resources. The databases on the server, and guides covering the API, transactions, authoring and publishing. An agent new to EvidentSource should start with evidentsource://guides/introduction.

Most responses come back as TOON, a compact form that includes links and the actions available next, so an agent explores by following them. A State View comes back in its own format.

See MCP API for the full list.

Why agents fit here

The two-function model gives an agent a narrow, checkable job: decide turns a command into events, and evolve folds events into state. And the agent’s actions are recorded like any other caller’s, as transactions in the log.

Your own deployment

Run the MCP server alongside your own EvidentSource. See MCP API.