CapForge MCP Adapter

ACM-controlled AI agent tool execution via the Model Context Protocol.

Overview

capforge-mcp is the official Model Context Protocol integration for CapForge. It lets you protect every MCP tool invocation against an Agent Capability Manifest (ACM) — a signed policy document that defines what actions an agent is allowed to perform.

Every tool call passes through ACM.check() before execution. Unauthorized attempts are rejected with a structured MCPPermissionError.

Installation

pip install capforge-mcp

Requires Python 3.12+.

Quick Start

from capforge import ACM
from capforge_mcp import ACMProtectedMCP

# Load your signed capability manifest
acm = ACM.load("manifest.json")

# Create an ACM-protected MCP server
mcp = ACMProtectedMCP("MyServer", acm=acm)

# Register a protected tool
@mcp.tool(resource="email", action="send")
def send_email(to: str, body: str) -> str:
    return f"Sent to {to}: {body}"

# Run the server
mcp.run(transport="stdio")

API

Dependency

capforge-mcp requires the capforge core package (pip install capforge) and the MCP Python SDK (mcp>=1.28,<2). Both are installed automatically as dependencies.

License

Apache 2.0