Answering Questions from Your Knowledge Base with MCP: A Step-by-Step Playbook

Every company has a knowledge base that people stop searching. Policies, how-tos and project notes live in Notion or Confluence, and staff ask a colleague instead. Answering those questions usually falls to team leads, ops managers and whoever wrote the page.

MCP (Model Context Protocol) lets an AI app search your knowledge base, read the matching pages and answer with a link to the source. The AI only reaches pages the signed-in person can already open. Done well, it turns "ask Sam" into "ask the wiki, then check the cited page."

This playbook covers Notion's and Atlassian's official MCP servers. Tool names come from each vendor's docs, checked on 28 October 2026. eSEOspace did not install or test these servers. More workflows are in our MCP playbooks hub.

What you need

Pick the server for the tool your company documents in. You can connect both if you use both.

Notion describes its server as "a remote MCP server hosted by Notion" in its developer docs. After you approve OAuth, the client can read and update content you can access. Our guide to connecting Claude to Notion walks through the setup.

Atlassian's getting started guide gives the server URL as https://mcp.atlassian.com/v2/mcp. It says OAuth 2.1 "ensures all actions respect users' existing access controls and permissions." Atlassian documents install paths for Claude, ChatGPT, Cursor, VS Code and other clients.

For the client, any app that supports remote MCP servers with OAuth will do.

The workflow

1. Collect the real questions. Pull ten to twenty questions people asked in chat last month. These become your test set and show where the wiki is thin.
2. Search first. For Notion, the AI calls notion-search. For Confluence, it calls searchConfluence, which searches with CQL (Confluence Query Language), or Atlassian's beta search tool for natural-language search across Jira, Confluence and connected apps.
3. Read the full page, not the snippet. Notion's notion-fetch returns the page. Atlassian's getConfluenceContent reads any Confluence content type.
4. Check freshness. In Confluence, listConfluenceContentVersions lists version history, newest first. Ask the AI to report the last edit date with the answer. In Notion, Business and Enterprise plans can filter search by last-edited date, per Notion's supported tools page.
5. Read the comments. Open questions and corrections often live in comments. Notion has notion-get-comments and Atlassian has listConfluenceComments.
6. Answer with citations. The answer should quote or paraphrase the page and link it. If two pages disagree, the AI should show both and say so.
7. Flag gaps instead of guessing. When nothing matches, the AI says so. Log the question for the page owner.
8. Suggest fixes, with approval. Notion's notion-create-suggested-edit proposes an edit, and notion-create-comment adds a comment. Atlassian's createConfluenceComment adds a footer or inline comment. Comments and suggestions let the page owner decide, which is safer than direct edits.

Example prompts

• "Search Notion for our travel expense policy. Fetch the best match, answer 'Can I expense a hotel upgrade?' and link the page. If it isn't covered, say so."
• "Use searchConfluence to find pages in the HR space about parental leave. Read the top two and tell me if they agree. Include each page's last version date."
• "Read the comments on the 'Onboarding checklist' Notion page and list any unresolved questions."
• "Here are 15 questions staff asked last month. For each, search Confluence and tell me whether a page answers it, with the link, or mark it as a gap."
• "Add a Confluence comment on the VPN setup page noting that step 4 refers to the old client. Show me the comment text before posting."

Guardrails

Start read-only. Both servers include tools that change content. Notion's list includes notion-update-page, notion-move-pages and delete tools that move items to Trash. Atlassian's write group includes tools that change page permissions and enable public links. In Confluence, ask your admin to grant only read_confluence and search_confluence for the pilot.

Keep approvals on for any write call. Comments and suggested edits are low risk. Page updates, moves, permission changes and public links should stay with a person.

Mind the reach. The AI can see everything the connected user can see. Notion's notion-connector-search can also search connected sources such as Slack and Google Drive, on Business or Enterprise with Notion AI. Atlassian's search can reach connected apps through Rovo. Read our pages on excessive permissions and cross-server data leakage, and see the MCP Security Center.

Use admin controls. Notion workspace owners manage client access in settings, and restricting which clients members can connect needs the Enterprise plan. Organization owners can list and revoke MCP connections through Notion's Admin API. Atlassian recommends least privilege and monitoring audit logs for unusual activity.

Keep these human: HR, legal and medical questions where a wrong answer causes harm, and any change to who can see a page.

What to measure

• Share of test questions answered with a correct, cited page.
• Share of questions the AI marked as gaps, and how many gaps were then written up.
• How often people still ask a colleague after asking the AI.
• Pages flagged as stale or contradictory, and how fast owners fixed them.
• Rovo credit use for Atlassian. Atlassian says usage is "per call" and depends on context volume and reasoning depth.

Next step

A knowledge base assistant is only as good as the pages behind it. If you want help connecting it safely and cleaning up the content it relies on, talk to us about Claude Code consulting.

Frequently asked questions