Agentic / AI
Every braised build produces a set of outputs for two distinct audiences: human readers navigating HTML in a browser, and agents (AI coding assistants, LLM pipelines, automated tools) consuming documentation programmatically. Both are first-class. Writers do not maintain separate content for each.
What braised produces
| Output | Location | Purpose |
|---|---|---|
| Companion Markdown | dist/*/index.md |
One per page. Agents can read every page without parsing HTML. |
llms.txt |
dist/llms.txt |
Site index for agents. Lets agents discover the site's content map without crawling. |
manifest.jsonl |
dist/manifest.jsonl |
Semantic chunks of every page. Feed this to your embedding pipeline. |
sitemap.xml |
dist/sitemap.xml |
Standard XML sitemap for search engines. Generated when site.url is set. |
| Artifacts | artifacts/ |
Structured JSON for deterministic lookups — nav tree, page metadata, CSS variables. |
Artifacts power the built-in MCP tools. manifest.jsonl is the interface for teams building their own search or RAG infrastructure — braised produces the chunks; the embedding model, vector store, and query layer are your choices.
Two roles
Writers control how their content appears to agents. The frontmatter fields llm_description and llm_exclude let writers override the default description or suppress a page from all agent outputs entirely. braised inspect --agent shows exactly what an agent sees for any page before the site goes live. See Authoring for Agents.
Operators deploy the tools agents connect to. braised mcp serve starts a long-running Streamable HTTP server that exposes built artifacts as structured MCP tools — nav tree, page index, page content, and CSS variables. See MCP Server for what it exposes and how it works, and Deploying mcp serve for production setup.
Two MCP modes
braised mcp serve (covered in this section) exposes your built site as queryable tools for agents — navigation, page content, metadata.
braised mcp (no subcommand) is a different tool: a stdio server that exposes braised's own schema — block syntax, braised.yaml format, nav rules — as MCP resources for AI authoring assistants like Claude Desktop and Cursor. It helps you write braised documentation, not query it. See CLI Reference for setup.