AI for your team wiki that shows its sources and waits for your approval.

AI search with sources

Ask a question in your own words. BookHost finds the matching passages in your workspace, writes a short answer only from those passages and lists every page and section it used, with a link. When nothing in the wiki fits, it says so instead of guessing.

• Available to workspace owners and admins, in the BookHost dashboard under Ask your wiki.
• Matches by words out of the box. Matching by meaning (beta) finds pages even when the question uses different words; an owner or admin turns it on with one click for workspaces without personal data, and turning it off deletes the index.
• Answers for every member, limited to what that member may read in BookStack, are planned. Until then, asking is restricted to owners and admins, who can already read every page.

Connect your AI agents (MCP)

Every workspace has its own MCP server at https://<your-workspace>.bookhost.co/mcp. Claude Code, Cursor, VS Code (GitHub Copilot agent mode) and the Codex CLI connect directly; Claude Desktop connects through the open-source mcp-remote bridge. Agents can search, read pages and attachments and, if you allow it, propose or make edits.

• Its own BookStack user. Owners and admins create an agent in the dashboard and pick its BookStack role; your existing book, chapter and page permissions decide what it sees.
• Propose only by default. Agent edits arrive in the review queue, marked as agent proposals. Nothing changes in the wiki until an owner or admin approves. You can switch a workspace to read-only or allow direct edits, which then show up in the revision history under the agent’s name.
• Kill switch and activity log. One button revokes every agent token; the log records tool, target and result, never page content or search terms.
• No extra AI processor. For this feature BookHost sends your pages only to the agent you connect, not to any model of ours.

Two-minute setup with Claude Code

1. Dashboard → Agents → create an agent and copy its token (shown once).
2. Add the server:

export BOOKHOST_TOKEN="<token id>:<token secret>"
claude mcp add --transport http --scope user bookhost https://your-team.bookhost.co/mcp \
  --header "Authorization: Bearer $BOOKHOST_TOKEN"

Then ask: “List the books in our wiki.” Configurations for Cursor, VS Code, Codex and Claude Desktop are on the agent setup guide. claude.ai and ChatGPT web connectors need OAuth sign-in, which is planned.

A real run against our showcase workspace:

$ claude "Using only the bookhost MCP server: list the books in our wiki,
  then search for 'onboarding' and tell me the top page result and its URL."

Books:
- Northwind Client Projects
- Northwind Studio Handbook
- Team handbook

Top "onboarding" page result: First week at Northwind
https://showcase-northwind.bookhost.co/books/northwind-studio-handbook/page/first-week-at-northwind

AI drafts with approval

Upload a PDF, DOCX, Markdown or TXT file and choose a destination book. BookHost drafts a wiki page from it with a summary, tags and a reviewer checklist. Your team edits the draft; only an owner or admin can publish it to BookStack, where the book’s permissions apply. AI output can contain mistakes, which is exactly why nothing is published without a person.

Sending documents by email is planned and not available yet.

Where your data goes

• Your wiki, its database and backups run on Hetzner infrastructure in Germany or Finland.
• AI search, AI drafts and matching by meaning use TensorX Limited (Dublin, Ireland); its documented GPU infrastructure is in Dublin and Helsinki, and its data processing agreement provides zero retention of prompts and completions. No other AI provider and no free-model fallback is used.
• Only what a feature needs is sent: the extracted text of an uploaded document, or a question and the passages it is answered from. Matching by meaning, when switched on, sends page text to compute search vectors (refreshed hourly), which are stored in our own EU database.
• Nothing goes to TensorX unless someone in your workspace uses one of these features. Agent access over MCP involves no AI provider of ours at all.
• You stay the controller. Details are in our privacy notice and data processing agreement (annex 3 lists TensorX).

Limits

• 100 AI answers during the trial, 1000 per calendar month (UTC) on Team, shared by the workspace. Questions that fail are not counted.
• 20 AI drafts during the trial, 300 per calendar month (UTC) on Team.
• Up to 20 questions per person per minute and 200 per workspace per hour.
• Agent tokens: 120 requests per minute, 1,500 per hour and 60 writes per hour each; up to 20 active agents per workspace.
• Uploads up to 10 MB and 60,000 extracted characters; scanned PDFs need OCR first.

AI is included in the €39/month plan; there is no separate AI add-on.

Beta and planned