To connect Confluence to an AI agent, ingest Confluence pages into Answer Engine as source-linked documents, then expose that library through npx answer-engine-mcp. The agent does not need a separate Confluence plugin; it calls the Answer Engine MCP server and retrieves scoped Confluence memory with page, space, version, and source metadata.
Last updated: June 20, 2026.
What is the shortest working Confluence-to-agent path?
The shortest path has three pieces: Confluence as the source, Answer Engine as the memory layer, and MCP as the agent interface. Atlassian documents Confluence Cloud REST API v2 page resources, including page bodies and body-format handling, in the official page API reference (Confluence REST API v2 pages). Answer Engine’s Confluence adapter is built around that shape: a page has an id, title, status, optional storage or view body, space metadata, version metadata, labels, ancestors, and links.
Once the page is ingested, the agent should not scrape Confluence again from the prompt. It should ask Answer Engine memory for the relevant page or answer. MCP is the portability layer here: the Model Context Protocol describes a client-server pattern where AI applications connect to external systems through a common protocol (MCP introduction). The same Answer Engine memory library can be represented in Claude Code, Cursor, or another MCP client without rebuilding the Confluence ingestion path.
Which adapter makes this honest?
The real adapter is src/services/content/adapters/confluence.adapter.ts. It validates Confluence page payloads and adapts them into Answer Engine content with contentType: "document". The adapter keeps useful page metadata:
| Confluence field | Answer Engine memory use |
|---|---|
space.key and space.name | Lets the agent distinguish engineering docs from support docs. |
version.number and version.when | Helps humans inspect whether a retrieved page is current enough. |
| labels | Become memory tags for filtering and review. |
_links.base + _links.webui | Preserves a source URL for cited answers. |
| ancestors | Keeps hierarchy context for nested pages. |
The Atlassian side is also real: src/services/integrations/atlassian.service.ts owns the Atlassian OAuth and sync paths. When you describe this publicly, say “Confluence pages become Answer Engine document memory.” Do not claim a generic wiki connector if the code path is specifically Confluence page ingestion.
What should you connect before MCP?
Start in the product, not the agent. Authorize Atlassian with the Confluence scope your workspace allows, then select the spaces or pages that should become memory. Atlassian documents OAuth 2.0 3LO app authorization separately from the Confluence page API, so treat authorization and page retrieval as two layers (Atlassian OAuth 2.0 3LO).
Keep the first sync narrow. A space with onboarding pages, runbooks, architecture notes, and support FAQs is a better first memory library than every historical page in the company. The goal is a grounded answer your agent can verify, not a giant unreviewed dump.
What MCP config should the agent use?
After the Confluence pages are in a library, configure the agent with the same Answer Engine MCP server used by the install guides. Replace the key and library id; keep the host fixed.
{
"mcpServers": {
"answer-engine": {
"command": "npx",
"args": ["answer-engine-mcp"],
"env": {
"ANSWER_ENGINE_API_KEY": "ae_live_your_key_here",
"ANSWER_ENGINE_API_URL": "https://engine.answeragent.ai",
"ANSWER_ENGINE_LIBRARY": "your-library-id"
}
}
}
}This config runs npx answer-engine-mcp and points the MCP server at the hosted Answer Engine API.
If you need the client-specific setup, use Add persistent memory to Claude Code or Add persistent memory to Cursor. Those guides show the same host and command in the right config file for each client.
What output should you expect?
After the MCP client starts the server, you should see a connected server named answer-engine:
answer-engine
status: connected
command: npx answer-engine-mcp
api: https://engine.answeragent.ai
library: your-confluence-library
tools: search_content, get_content, ask, save_content, append_memoryThen ask a question that only a Confluence page can answer:
Use Answer Engine memory to find the deployment freeze process from Confluence. Cite the page you used.A good result names the process and points back to the Confluence page source. If the answer has no source, treat that as a wiring failure. You are checking source-aware recall, not trying to publish a benchmark.
How should you structure Confluence memory?
Create libraries that match agent jobs. A support agent might need customer-facing FAQs, incident runbooks, and escalation rules. An engineering agent might need architecture decisions, service ownership, and deployment docs. Mixing all spaces into one library can work later, but it makes first-run verification harder.
For a broader rubric, read Agent Memory: The Complete Guide. The important properties are source lineage, tenant isolation, supersession handling, and inspectability. Confluence gives the source pages; Answer Engine keeps those pages callable through MCP.
Which sources ground this setup?
- Confluence REST API v2 pages
- Atlassian OAuth 2.0 3LO
- Model Context Protocol introduction
- MCP architecture
FAQ
Can an AI agent use Confluence pages through MCP?
Yes. The agent connects to Answer Engine through MCP, and Answer Engine retrieves Confluence pages that were ingested into the selected memory library.
Does Answer Engine store Confluence as a special content type?
The Confluence adapter adapts pages into document content while preserving Confluence-specific metadata such as page id, space key, version, labels, hierarchy, and source URL.
Should I paste Confluence text directly into the prompt?
Use direct prompting only for one-off experiments. For repeated agent work, ingest the page once and let the agent call memory through MCP so the source trail remains inspectable.
What should I test first?
Pick one page with an answer you can recognize, ask the agent to use Answer Engine memory, and verify that the answer cites the Confluence source page.