forge-mcp

MCP Meta-Orchestrator for Forge — routes, secures, budgets, and caches tool calls across N upstream MCP servers.

Features

• Deny-by-default security: tools must be explicitly allowed: true per server config
• Caller allowlists: restrict tool access to specific agent/caller IDs
• Secret redaction: responses are scanned and sanitized before reaching the agent context
• Per-run budgets: configurable max_calls_per_run per tool; BudgetExceededError on breach
• Exact-match TTL cache: Anthropic prompt-caching semantics — identical input → cache hit
• Circuit breakers: per-server, reusing forge-core.CircuitBreaker (5 failures → OPEN)
• YAML config: human-friendly server and policy authoring
• CLI: forge mcp list-tools, forge mcp call, forge mcp status

Quick start

from forge_mcp import MetaMCPServer, ServerConfig, ToolCall, ToolPolicy, FakeMCPClient

server = MetaMCPServer()
cfg = ServerConfig(
    name="filesystem",
    transport="stdio",
    command=["npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
    tools={"read_file": ToolPolicy(allowed=True, max_calls_per_run=20)},
)
server.add_server(cfg, FakeMCPClient(responses={"read_file": "file contents"}))

result = await server.call_tool(ToolCall(tool_name="read_file", arguments={"path": "/tmp/x.txt"}))
print(result.content)

YAML config

cache_ttl: 300
servers:
  - name: filesystem
    transport: stdio
    command: ["npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
    tools:
      read_file:
        allowed: true
        max_calls_per_run: 20
      write_file:
        allowed: true
        max_calls_per_run: 5

Part of Forge

This package is part of Forge — the universal AI agent harness.