Sources
Two sources are there before you add any:
- Your site's own repository — the one your docs are built from, listed first and never removable.
- Your product website — the address in Settings ▸ General ▸ Product website, read for real prices, plans and limits.
What can be a source?#
The address decides the kind, whichever card you paste it into:
| You paste | It becomes | A read returns |
|---|---|---|
github.com/acme/api |
Repository | Its readable files, README and docs first; any file by path; the last 10 commits |
github.com/acme/api/tree/main/docs |
Repository folder | Only that folder |
acme.com or acme.com/docs |
Website | Up to 10 pages at a time, found from its sitemap.xml |
acme.com/pricing.html |
Single page | That one page, read whole |
A file extension in the path makes it a single page. Tracking parameters such as utm_source are dropped, so the same page pasted twice is one source.
In Integrations, the Sources filter lists the named kinds too, each read the same way:
- Documentation platforms — Mintlify, GitBook, ReadMe, Docusaurus, Read the Docs, MkDocs, Redocly, Scalar and more.
- Knowledge bases — a public Notion page, Zendesk Guide.
- Code and APIs — an OpenAPI spec, an npm or PyPI package page.
- Community — a Discourse forum's public topics.
- Files — upload a PDF, Word file, Markdown or plain text; each becomes a page of your docs.
Cards Docsbook cannot read yet — a Notion workspace, Confluence, GitLab, Slack and others — are drawn muted and say what they would need.
Connect a source#
Connect from the panel, or ask for it from your editor.
In Integrations, open Discover, pick the Sources filter, and open the card for what you have — GitHub repository, Website, OpenAPI spec, Mintlify.
Press Add addresses and paste as many as you like, one per line. The dialog shows which card each address will land on before you commit.
Fill in Why this matters, in your words — for example, "The prices here are the only correct ones." The agent reads it as an instruction, and it applies to every address in the batch.
Press the Connect button at the foot of the dialog. Each address reports back on its own line, and the card lists it under Addresses, marked "not read yet" until the agent reads it.
From Claude Code, Cursor or Codex, your MCP connection does the same with connect_source. It proves a repository is readable before storing it, and needs a token authorised with Allow editing documentation.
Private repository? Install the Docsbook GitHub App on it first — Contents: read is enough. Reads go through that installation, so they also work in scheduled runs with nobody signed in; without it, GitHub answers as if the repository did not exist.
What the agent does with a source#
Before it writes or judges a page, the agent lists your sources and reads the ones that bear on the task. Your note travels with each source as an instruction.
These trigger cards run on sources by themselves. The ones that wake when your docs "finished re-indexing" need search by meaning on, because that re-index is what wakes them.
| Card | Wakes | What it does |
|---|---|---|
| Find pages the code outgrew | Your docs changed and finished re-indexing | Reads what changed in the source, fixes what the code plainly contradicts, proposes the rest |
| MCP sync | Daily | Compares every tool your MCP server registers with what the docs say about it |
| OpenAPI sync | Daily | Rebuilds the API reference from your spec |
| SDK sync | Daily | Keeps the reference in step with what your SDK exports |
| Document the API surface | Your docs changed and finished re-indexing | Finds endpoints, options and errors that exist in code but nowhere in the docs |
| Write the release notes | Your docs changed and finished re-indexing | Turns the commits and merged work of a release into notes |
OpenAPI sync ships pointing at a demo spec. Replace that link in the card's prompt with your own spec's address — until you do, it changes nothing and reports that the link needs replacing.
To wake the agent when code lands rather than when docs change, connect the GitHub card in Integrations and arm an occasion: Code landed and API spec changed are checked every 15 minutes, Release published hourly. See Triggers.
How much of a source is read?#
A read happens when the agent asks for one, against the live address:
- Repository listing — up to 300 files, README and docs first; dependency and build folders are left out of the list, but any file can be read by its path.
- One file or page — up to 25,000 characters, marked when cut.
- A website without a path — up to 10 pages from its sitemap, 8,000 characters each; with no sitemap, only the page you connected.
- Commits — the last 10: message, author and date.
Change or remove a source#
Ask the agent, or call configure_source from your editor with a read-write token. It renames a source, rewrites its note, pauses it (enabled: false) or removes it (disconnect: true) along with any GitHub authorisation attached to it. See the MCP tools reference.
FAQ#
Does the AI chat on my site read my sources? No. It answers readers from your published pages; sources are read by the agent, the admin chat and your own MCP client.
Does Docsbook copy my sources? No. Nothing is mirrored or crawled on a timer: a source is read when the agent needs it, and website reads respect robots.txt.
Why does a website read return only one page? Its section is not in the site's sitemap.xml, or there is no sitemap. Ask for a specific path, or connect the repository behind the site instead.