Using MCP Servers to Extend Your AI Agent

What MCP Is

Model Context Protocol (MCP) is a standard for connecting AI agents to external tools. An MCP server exposes tools over a stdio or HTTP transport. Your agent calls these tools the same way it calls built-in tools.

The protocol is open. Any tool can ship an MCP server. Many do.

Adding an MCP Server to Hermes

Most MCP servers ship as npm packages. Add one to your Hermes config:

hermes tools add \
  --name firecrawl \
  --command "npx -y firecrawl-mcp" \
  --env FIRECRAWL_API_KEY=your_key

Or edit the config directly:

1
mcpServers:
2
  firecrawl:
3
    command: npx
4
    args: ["-y", "firecrawl-mcp"]
5
    env:
6
      FIRECRAWL_API_KEY: "your_key"

Restart Hermes after adding a server.

Verifying the Tool Is Available

hermes tools list

You should see the MCP server's tools listed. For firecrawl, that is `firecrawl_scrape`, `firecrawl_search`, and others.

Test it in a session:

Use firecrawl to scrape https://example.com and return the page title.

The agent will call `firecrawl_scrape` and return the result. If the tool does not show up, check the server logs:

hermes tools logs firecrawl

Common MCP Servers Worth Adding

• `firecrawl-mcp` for web scraping
• `@notionhq/notion-mcp-server` for Notion read/write
• `n8n-mcp` for triggering n8n workflows
• `figma-developer-mcp` for reading Figma designs

Pitfalls

• MCP servers start as child processes. If the npm package is not installed, the server will not start.
• Environment variables must be set in the MCP server config, not just in your shell.
• Tool names in MCP are namespaced by server. `firecrawl_scrape` is different from a local tool called `scrape`.