MCP for AI agents

EvidentSource exposes an MCP (Model Context Protocol) endpoint. AI agents — Claude, GPT-based agents, custom LLM harnesses — can reach the Sandbox directly and:

• Discover available commands, state views, and databases
• Issue commands
• Read state views (with bi-temporal parameters)
• Query event streams and decision traces

Endpoint

https://sandbox.evidentsource.com/mcp

No authentication for the public Sandbox. (See About the Sandbox.)

Why MCP matters here

The two-function programming model gives LLMs the right shape of problem — pure functions are easy targets for code generation. MCP closes the loop: the agent can both read the system’s decision traces (to inform its next action) and issue commands (to act). Every event it produces is captured as a decision trace, making the agent auditable by default.

In an agent config

For Claude Desktop or a Claude Code plugin, add the Sandbox MCP endpoint to your MCP configuration. Refer to modelcontextprotocol.io for the exact config format.

Tools the MCP endpoint exposes

• list_databases
• list_commands(database)
• issue_command(database, command)
• list_state_views(database)
• read_state_view(database, view, params, as_of?)
• query_events(database, stream?, subject?, type?, as_of?)

Production MCP

Your own EvidentSource deployment can expose MCP the same way, gated by your own authentication. See Security.