mcp

Dieser Inhalt ist noch nicht in deiner Sprache verfügbar.

MCP (Model Context Protocol) is mashin’s implicit surface. Every machine is automatically available as an MCP tool with zero configuration. When you write a machine and start the MCP server, AI assistants like Claude, Cursor, ChatGPT, and VS Code can discover and call it.

No expresses declaration is required for MCP. It is always on.

Overview

MCP is the default way machines are consumed. The mashin MCP server implements the Model Context Protocol over stdio (JSON-RPC 2.0). It exposes every machine in the current cell as a callable tool, with input schemas derived from accepts and output schemas from responds with.

Because MCP is implicit, the expresses section is only needed when you want surfaces beyond MCP, such as REST APIs, webhooks, or web pages.

Syntax

MCP requires no syntax. Every machine is MCP-accessible by default:

machine classify_email

  accepts
    subject as text, is required
    body as text, is required

  responds with
    category as text
    priority as text

  implements
    ask classify, using: "anthropic:claude-sonnet-4-6"
      with task "Classify this email by category and priority"
      returns
        category as text
        priority as text
      assuming
        category: "general"
        priority: "normal"

Start the server with mashin mcp. The machine appears as a tool named classify_email.

Configuration options

MCP has no per-machine configuration. It is controlled at the cell level:

Setting	Scope	Description
Trust level	Cell	System (local stdio), Authenticated (API key), or Untrusted (no auth)
Rate limiting	Cell	Per-organization, per-minute request limits

Local connections from mashin mcp run at system trust. Remote MCP endpoints require authentication.

Tool naming

MCP tools are named after the machine. Multi-chain machines expose one tool per chain:

Machine	MCP tool name
`classify_email`	`classify_email`
`data_pipeline` (single chain)	`data_pipeline`
`multi_tool` with chains `search`, `count`	`multi_tool/search`, `multi_tool/count`

Input schema generation

The accepts section maps directly to the MCP tool’s JSON Schema:

accepts
  query as text, is required
  max_results as number
  include_archived as boolean

Produces:

{
  "type": "object",
  "properties": {
    "query": {"type": "string"},
    "max_results": {"type": "number"},
    "include_archived": {"type": "boolean"}
  },
  "required": ["query"]
}

Governance

Every tools/call invocation goes through the full governance pipeline:

Governance permissions from ensures are checked
Inputs are validated against accepts
The machine executes
A SurfaceAccess event is recorded with :mcp surface in the behavioral ledger
Budget limits are enforced

Denied calls return a structured error. The governance decision and reason are recorded in the ledger regardless of outcome.

Example

Connecting to Claude Desktop:

{
  "mcpServers": {
    "mashin": {
      "command": "mashin",
      "args": ["mcp"]
    }
  }
}

After restarting Claude Desktop, all machines in your cell appear as tools. Ask Claude to classify an email and it calls the classify_email tool, which runs through governance, executes, and returns the result.

mcp