arc-gate-mcp 0.1.1

Runtime governance for MCP tool calls — Arc Gate for the MCP protocol layer

arc-gate-mcp

Runtime governance for MCP tool calls.

Arc Gate MCP sits between your agent and any MCP server. It intercepts all tool call results and enforces instruction-authority boundaries before the agent processes them.

When a tool result contains injected instructions — a poisoned document, a malicious webpage, a hostile database row — Arc Gate blocks them before they reach the agent.

Install

pip install arc-gate-mcp

Usage

Full proxy (wraps any MCP server)

from arc_gate_mcp import ArcGateMCPProxy

proxy = ArcGateMCPProxy(
    upstream_url="http://localhost:8000/sse",
    policy_mode="rag_assistant",
)
proxy.run()

Per-tool guard

from arc_gate_mcp import ArcGateToolGuard

guard = ArcGateToolGuard(policy_mode="rag_assistant")

@mcp.tool()
async def read_document(path: str) -> str:
    content = read_file(path)
    return guard.check(content, tool_name="read_document")

CLI

arc-gate-mcp --upstream http://localhost:8000/sse --policy rag_assistant

Policy modes

Mode	Behavior
balanced	Block on detected injection
browser_agent	Strip injections, allow safe content
finance_agent	Strictest — block everything suspicious
rag_assistant	Strip injections, preserve safe data

License

AGPL-3.0. Commercial license available — contact [e-mail].

Related

• arc-gate — Runtime governance proxy for LLM agents
• arc-sentry — Whitebox detector for self-hosted models
• arc-gate-mcp — Runtime governance for MCP tool calls
• arc-gate-benchmark — Reproducible benchmark harness