Core Concepts
AI-Readable Artifacts
Learn about the structure and use cases of AI-readable artifacts.
During build, Anydocs generates outputs for reader search and outputs for machine consumption. Knowing the difference tells you what humans should use and what AI should read.
Separate search outputs from AI outputs first
Not every generated JSON file is a good primary AI input.
- `search-index.<lang>.json`: powers reader search and helps humans find pages and sections
- `llms.txt`: lightweight AI discovery entry for the site structure
- `llms-full.txt`: full-site sequential text export used as a coarse fallback
- `mcp/*.json`: structured machine-readable outputs for targeted agent retrieval
Text-based AI entry points
When an external agent only needs fast discovery or coarse full-site reading, start with the text outputs.
- `llms.txt`: lighter and better for discovering the site and page entry points
- `llms-full.txt`: heavier and better for coarse full-site retrieval
- Both files are build outputs and contain only `published` content
Structured machine-readable outputs
When an agent needs precise retrieval by page, navigation, or content chunk, prefer the `mcp/` outputs.
- `mcp/index.json`: artifact index
- `mcp/pages.<lang>.json`: page-level summaries and public metadata
- `mcp/chunks.<lang>.json`: chunked content, usually the best fit for targeted retrieval
- `mcp/navigation.<lang>.json`: published navigation data
- `mcp/openapi/index.<lang>.json`: machine-readable API source index when configured
Public visibility and metadata exposure
AI outputs do not expose everything that exists in the source project.
- Only `published` pages enter `llms.txt`, `llms-full.txt`, and `mcp/`
- `draft` and `in_review` pages remain in the canonical source but do not enter public AI outputs
- `mcp/pages.<lang>.json` only exposes metadata whose visibility is `public`
Recommended read order for agents: A common safe pattern is: start with `llms.txt` for discovery, then switch to `mcp/navigation.<lang>.json`, `mcp/pages.<lang>.json`, and `mcp/chunks.<lang>.json` for targeted retrieval instead of loading the whole site at once.