How does the Traverse MCP server work?

The Traverse MCP server is a stdio process that implements the Model Context Protocol. It sits between your AI agent (Claude, GPT, or any MCP-compatible client) and your registered capabilities. The agent asks what tools exist, the server returns them, and the agent calls them by name.

The two core methods

tools/list — returns all capabilities currently in the registry as MCP tool definitions. Each tool definition includes the capability name, a description drawn from the contract, and the JSON schema for the input. The agent uses this to decide which tool to call.

tools/call — takes a capability name and a JSON input, runs it through the full Traverse runtime (precondition check, WASM execution, postcondition check), and returns the output. If any validation fails, the error goes back to the agent with a structured message.

Starting the server

traverse mcp serve

This starts the stdio server on the current registry. To connect it to Claude Desktop, add it to your MCP config:

{
  "mcpServers": {
    "traverse": {
      "command": "traverse",
      "args": ["mcp", "serve"]
    }
  }
}

Why this matters for AI agents

Without contracts, an agent calling a tool is guessing. It sends inputs it thinks are valid and hopes the output matches what it expected. With Traverse, the agent gets the schema upfront. The runtime validates the call before execution. The agent gets a typed, predictable result every time.

See how AI agents discover capabilities for more on the discovery flow.