arc-gate-mcp is a runtime governance proxy for MCP (Model Context Protocol) tool calls, published on GitHub by 9hannahnine-jpg. Its core positioning is as a security gateway at the MCP protocol layer, establishing a governance barrier between MCP clients and any MCP server.
In terms of purpose, arc-gate-mcp mainly addresses the problem of MCP trust boundaries being exploited. When an AI agent calls external tools via MCP, the content returned by the tools may contain maliciously crafted prompt injection instructions, or induce the agent to perform operations beyond its permissions. arc-gate-mcp performs governance checks on tool results before they reach the agent, thereby blocking such attacks and permission abuse.
In terms of main capabilities, the tool provides multiple governance policies to choose from: balanced is suitable for general scenarios; browser_agent is aimed at web browsing agents; finance_agent is aimed at financial data agents; rag_assistant is aimed at document retrieval agents; and strict provides the maximum strength of enforcement. These policy names indicate that it has differentiated configurations for different risk scenarios.
In terms of usage, arc-gate-mcp is deployed as a proxy between MCP clients and MCP servers, and all tool results flow through the Arc Gate governance layer. The material mentions that it can be used with Claude Desktop, but does not provide specific installation commands or configuration steps. In addition, arc-gate-mcp has been launched as a template on the Heym enterprise agent platform, and users can experience its actual operation through the Governed Web Research Agent template.
Applicable scenarios include: agent applications that need security review of content returned by external MCP tools, web browsing agents, financial data agents, document retrieval agents, and any MCP integration environment with protection needs against prompt injection and permission abuse.
In terms of limitations, the available material does not explain the MCP versions it supports, deployment dependencies, performance overhead, whether it is open source and commercially usable, specific installation commands, or configuration details; this information is not currently provided.
