Governed Web Research Agent (MCP)

Governed Web Research Agent (MCP)

This template shows an Agent that fetches and summarizes web pages, with every fetched page routed through a governance proxy before it reaches the model. The Agent's Fetch MCP server (uvx mcp-server-fetch) runs behind Arc Gate MCP (uvx arc-gate-mcp), a runtime proxy that screens tool results for prompt-injection and capability abuse.

What this workflow does

1. ResearchQuestion provides a URL or research prompt
2. WebResearchAgent calls the Fetch MCP tool; the page passes through Arc Gate's browser_agent policy first
3. The Agent summarizes the governed content and notes anything Arc Gate flags or blocks
4. ResearchSummary returns the final summary

Use cases

• Summarizing untrusted web pages where fetched content could hide instructions
• Research or monitoring agents that read arbitrary URLs
• A reusable pattern for putting a governance proxy in front of any MCP tool

Setup

1. Get a free Arc Gate API key at bendexgeometry.com and add it to the Agent's MCP connection env.
2. Connect an LLM credential in the Agent node and keep the policy on browser_agent.

The Agent's MCP connection is preconfigured (uvx arc-gate-mcp wrapping uvx mcp-server-fetch); Heym launches it on run, so there is no manual install step. The PyPI package is at pypi.org/project/arc-gate-mcp.

Notes

Arc Gate MCP is a third-party governance proxy and was released recently, so review its policy and data handling before production use - it proxies all fetched content and uses an external API key. To run the Fetch MCP directly without governance, point the Agent's MCP connection at uvx mcp-server-fetch.