Sources
Open the Sources section of your project's admin panel, directly under MCP and Prompts.
The catalogue#
The section is a table of every kind of source Docsbook knows about, grouped as Your project, Documentation platforms, Knowledge bases, Code and APIs and Community. Rows you have connected are lit; the rest are grey, and most of them connect in one paste — a published documentation site is a website whoever built it, so Mintlify, GitBook, ReadMe, Docusaurus, Read the Docs and the others need nothing set up per vendor.
Grey means two different things, and the row says which:
- Not connected — the row offers a Connect. Reading it works today.
- Not available yet — the row offers no button, and prints what it would need: an authorisation you grant once, a bot token, or a reader for a repository host we do not read yet. Where a workaround exists the row names it — a public help centre connects as a Website today.
A row without a button is deliberate. A Connect that cannot connect would make every other row on the screen untrustworthy.
Connecting one#
Press Connect on a row, or the New Source button above the table for anything at all. Either way there is one field, and the address decides what the source is and which row it lands under — not the row you pressed:
| What you paste | What it becomes | What gets read |
|---|---|---|
github.com/acme/api |
Repository | Its files, on demand — README and docs first |
github.com/acme/api/tree/main/docs |
Repository folder | Only that subtree |
acme.com or acme.com/docs |
Website | Several of its pages, found from its own sitemap.xml |
acme.com/pricing.html |
Page | That one page |
acme.mintlify.app, acme.gitbook.io, acme.notion.site… |
That vendor's row | The same as a website, filed where you would look for it |
If the address disagrees with the row you opened, the dialog says so before you commit.
A link pasted from an issue or a pull request connects the repository, not the issue. Tracking parameters (?utm_source=…) are dropped, so the same page pasted from a tweet and from your address bar is one source rather than two.
A project can connect more than one of the same kind — two websites, two repositories, two forums. Use New Source, or + Add on a row that already has one.
Two entries appear without you adding them, and neither can be removed here:
- This site's repo — the repository your documentation is built from.
- From Branding — the Site source URL, if your workspace set one. It still lives on the Branding card.
Each connected source can carry a short note about why it is connected. That note is not decoration: the assistant reads it as instruction.
Live, paused, and what the green dot means#
Every connected source shows a green dot and the word LiveSync.
That badge means the source is connected and your agents may read it. It is not a health check. Nothing pings the host, and nothing checks the repository still exists — a status light that no code writes is worse than no light at all. The honest signal is Last read, shown on the row and written only when a tool actually fetched the source successfully.
A provider's connections show as chips in the last column; open the row to see each one's note, when it was last read, and its controls.
Pause keeps a source in the list and stops anything reading it. Remove disconnects it. The two are different on purpose: "stop reading this for now" should not make you retype the address later.
What reads them#
The assistant in your admin panel. Its two tools, list_sources and read_source, are available on every turn rather than behind a lookup, and its instructions name your connected sources literally — a rule that says "check your sources" without naming them is one a model cannot follow.
Your own MCP agents. The same two tools are served over your project's MCP endpoint, so a source you register means the same thing in Claude Code or Cursor as it does here. See the MCP server reference.
Background runs. Scheduled prompts and agent runs read them too, which is where it matters most: there is nobody sitting there to paste a link.
A repository read without a path lists its readable files; with one, it returns that file. A website read without a path returns several of its pages as Markdown, discovered from the site's own sitemap and scoped to the section you connected — connect acme.com/docs and the blog stays out of it. Sites with no sitemap return the entry page alone, and the result says so, so a thin read is never mistaken for a thin site.
Which runs can actually read them#
Not every tool reaches a source, so the panel says which do rather than implying they all can. Wherever a run is listed — the Sources column in the Prompts table, and the strip under an MCP tool's description — the chips are drawn in two states:
- Lit — this run fetches the source: your site's pages, your repository's files.
- Unlit — this run knows the source is connected and will not fetch it. Ask the assistant instead; it has no such limit.
- Nothing at all — this run reaches no source. A settings write has no business reading your repository, and a chip there would say otherwise.
What it is for#
Prices, plan names, limits, endpoints, version numbers and behaviour all change faster than documentation does. Without a source, an assistant asked about any of them either says it does not know or invents something plausible — and a plausible invention is the version that ships to your readers.
With one connected, that question has an address. The assistant reads your pricing page and corrects the three pages that quote the old tier; it reads the repository and finds the four things your quickstart claims that the code stopped doing.
Related#
- MCP Server — the same two tools, for your own agents
- AI Chat — the assistant on your docs site
- MCP tools reference —
list_sourcesandread_sourcein full