langchain-mcp-trust-gate

Trust-gated MCP tool calls for LangChain. Automatically checks behavioral trust scores via Dominion Observatory before every MCP tool call.

Install

pip install langchain-mcp-trust-gate

Quick Start

from langchain_mcp_trust_gate import TrustGateInterceptor
from langchain_mcp_adapters.client import MultiServerMCPClient

trust_gate = TrustGateInterceptor(min_score=70)

async with MultiServerMCPClient(
    connections={"my-server": {"url": "https://example.com/mcp", "transport": "streamable_http"}},
    interceptors=[trust_gate],
) as client:
    tools = await client.get_tools()
    # All tool calls now trust-gated automatically

How It Works

1. Before each MCP tool call, queries Dominion Observatory for the server trust score
2. Blocks calls to servers below the threshold (default: 60)
3. After each call, reports the outcome back to the Observatory
4. Scores are cached (5-minute TTL) to minimize latency

Configuration

TrustGateInterceptor(
    min_score=70,           # Minimum trust score (0-100)
    cache_ttl=300.0,        # Cache TTL in seconds
    block_on_error=False,   # Fail-open if Observatory unreachable
    report_outcomes=True,   # Report call outcomes back
)

Architecture

Implements the ToolCallInterceptor protocol from langchain-mcp-adapters. Composes with other interceptors in an onion pattern.

License

MIT