Skip to main content
Not built yet. This page describes what is planned. Today, agents reach an environment’s tools through the agent contract — a function Silo calls with callTool.
An MCP server would expose a Silo environment’s tools over the Model Context Protocol, so any MCP-capable client could act inside a simulated world without writing an agent module.

​
What it would unlock

Drive a simulated world from an existing MCP client. Point a client that already speaks MCP at a Silo environment and its 42 CRM tools become available as ordinary tools — no adapter, no glue. Try an environment by hand. Ask for a stalled deal, reassign it, see what the tools say, before writing any agent code. Today the fastest way to explore a tool surface is a scripted agent and a trace. Evaluate agents you did not write. Anything that speaks MCP could be pointed at an environment, which widens what Silo can measure beyond agents you can wrap in a function.

​
The open question

Silo’s value comes from a rollout being a bounded, graded unit: one fresh world, one task, one verdict, one set of artifacts. An MCP session is the opposite shape — a long-lived connection with no natural start or end. So the design question is not how to expose the tools. It is what a rollout means over a session:
  • When does the world get created, and when is it graded?
  • Does one connection map to one task, or can a client switch tasks?
  • What does a trace look like when the client, not Silo, decides when work is finished?
Those answers determine whether state-diff.json and result.json still mean anything. Getting it wrong would make MCP a way to poke at an environment without ever producing an evaluation — interesting, but not what Silo is for.

​
What works today

Write an agent function. It is a small contract, and a model loop in any framework fits it in a few lines:

Bring Your Own Agent

The full contract, tool errors as values, and where a model loop plugs in.
For exploring a tool surface by hand, a scripted agent plus trace.jsonl gives you the same information today — see Debug a Failed Run.