Use with Agents

Distinguish site artifacts, hosted docs MCP, and the CLI MCP — and run Bengal as an MCP server while you author a site

2 min read 398 words
Edit this page

Was this page helpful?

Bengal talks to agents in three places. Mixing them up is the usual failure: a published site'sllms.txtis not the CLI, and the Connect to IDE button is not bengal --mcp.

This page is for people authoring a Bengal site who want the CLI inside an MCP client. For files agents read on a published site, see AI-Native Output. For a button that installs a hosted docs MCP, see Connect to IDE.

Three surfaces

Surface Audience What Bengal provides
Site artifacts (llms.txt, agent.json) Agents reading a published site Generated on everybengal build
Hosted docs MCP Readers installing your docs in an IDE The Connect to IDE button. Bengal does not provide that HTTP server
CLI MCP People authoring a Bengal site bengal --mcp— JSON-RPC on stdin/stdout

Run the CLI as an MCP server

bengal must be on your PATH, the same as any other CLI.

BASH
bengal --mcp

That process speaks MCP over stdio (JSON-RPC on stdin/stdout). Point your MCP client at this command. It exposes the registered Bengal command tree — the same public CLI you already use, including bengal build and bengal serve.

This is not an MCP server for your published documentation. It does not replace a hosted Streamable HTTP endpoint, and it is not a Python entry-point plugin. For that extension surface, see Writing Plugins.

Register with the Milo gateway

To register this CLI in the Milo gateway for AI agent discovery:

BASH
bengal --mcp-install

To remove it:

BASH
bengal --mcp-uninstall

These flags talk to the Milo gateway only. They do not install a hosted docs MCP, and they do not change whatbengal build writes into public/.

CLI command catalog for agents

BASH
bengal --llms-txt

Prints an agent-readable catalog of the registered CLI. That stdout dump is not the site-levelllms.txt file produced by bengal build— see AI-Native Output for the published-site artifact.

See Also

  • AI-Native Output —llms.txt, agent.json, and other files for agents reading a published site
  • Connect to IDE — button for a hosted docs MCP (Bengal does not provide that server)
  • Writing Plugins — Python entry-point plugins; a different extension surface
  • Output Formats — JSON, LLM text, and other machine-readable outputs