Documentation Agents
Help users find answers in your published documentation through chat and MCP.
A documentation agent answers questions from your published docs, reads the surrounding code and instructions, and links readers to the pages that support its answer. Serve it in your documentation site, or expose its retrieval tools through MCP so developers can use the same documentation from their coding agents.
The Agno Docs Agent uses this pattern. Its application chooses when to retrieve evidence, how much context to include, and how to cite it. Agno provides page storage and search, a read-only page filesystem, and the AgentOS runtime.
Follow a question to its sources
- Search the current published pages for the user's question. For a follow-up, use the previous question to help resolve its context.
- Read the best matching pages at the revisions returned by search. Include complete pages where the context budget permits, then excerpts.
- Let the agent search again or inspect exact identifiers when the initial evidence is insufficient.
- Answer with links beside the claims they support. When retrieval is unavailable or the docs do not answer the question, say so.
This combines a predictable first retrieval with tools the agent can choose to use. It is useful for API questions where a short search excerpt might omit a required import, a parameter default, or a later setup step.
Share the documentation through chat and MCP
| Surface | Reader experience |
|---|---|
| Site chat | Ask a question, continue the conversation, and open cited pages |
| MCP | Give another assistant tools to search and read your documentation |
| Deployment-triggered sync | Update the published page index after a docs deployment |
MCP clients use their own models to reason over the tool results. The chat agent's instructions and citation rules belong to the chat application; they are not automatically applied to every MCP client.
Keep the index aligned with the site
Index the site's clean Markdown through llms.txt. This captures the published content, including resolved snippets, rather than asking retrieval to interpret the source site's MDX components.
Published Pages keeps page Markdown and its search vectors together. Each page is published atomically, and reads can request the revision found by search. The sync report distinguishes a completed reconciliation from a partial result.