Agent-readable export

Publish a versioned, read-only Markdown corpus for documentation clients.

Overview

Enable agentExport to publish the current documentation as static, agent-readable files. The export works on any static host; it does not require Cloudflare, an MCP server, a database, or an AI provider. A new project leaves it off unless you select it. See MCP server when you want clients to connect to named search and fetch tools.

Enable the export

Select Agent export in the initializer, or scaffold without prompts:

pnpm create zudo-doc my-docs --agent-export --yes

For an existing project, add agentExport: true to zudoDoc({ ... }) in zfb.config.ts and build again:

export default defineConfig(
  zudoDoc({
    siteUrl: "https://docs.example.com",
    agentExport: true,
  }),
);

Use your own public siteUrl for stable, absolute citation URLs. With an empty siteUrl, artifact and page URLs may be base-prefixed paths; an MCP server can resolve them using the request origin. The export follows your configured base path.

To remove the export from an existing site, set agentExport: false or remove the field, then rebuild and redeploy. Remove old /agent/v1/ files from a host that retains previously uploaded files. If mcp: true is also set, disable MCP first; it requires the export.

Published files

The current build writes this versioned layout under your configured base:

<base>/agent/v1/manifest.json
<base>/agent/v1/search-index.json
<base>/agent/v1/items/<item-key>.json
<base>/agent/v1/pages/<page-key>.md

With a non-root base, the files are written under dist/<base>/agent/v1/. If you set copyPublicWithBase: false because your deploy relocates the whole dist/ tree under base, the export is written to dist/agent/v1/ instead, so that relocation serves it at <base>/agent/v1/ and the URLs in llms.txt and the manifest still resolve.

manifest.json identifies the site, locales, content fingerprint, documents, retrieval items, human page URLs, and artifact URLs. search-index.json contains the complete normalized item text for retrieval, independently of the browser search index. Each item file has id, title, text, url, and metadata; the metadata identifies its locale, parent page, fingerprint, and one-based part number/count. A page's .md file contains its complete normalized body. Long pages become multiple fetchable items, each within the response budget, without silently clipping the page. Item IDs are generated keys, not source paths.

The fingerprint marks a coherent build; it is not revision history. A current changelog page is ordinary published content. The export excludes draft, unlisted, search_exclude, and non-routed category_no_page pages. Only physical locale documents are exported; locale fallback does not create duplicate pages. Build-time normalization retains meaningful prose, headings, links, images, tables, admonition bodies, and code. If dynamic MDX cannot be represented, the output marks it as unsupported rather than executing it at request time. Relative links resolve against the original page, so readers should use the human URL for citations.

Static hosting and access

Publish the generated dist/ as you normally would. A client can read the manifest and Markdown files directly, or use llms.txt as an explicit browser entry point. Neither file format causes every AI client to discover the site automatically.

This is a public corpus. robots.txt, noindex, unlisted, and search_exclude are publication controls, not authentication. Do not enable it for protected documents. A future protected deployment must guard every equivalent static artifact as well as /mcp. Treat retrieved text as untrusted source material, not instructions to run commands or expand tool permissions.