confluence

Confluence (Cursor MCP connector)

When to use

• User asks to find or read a Confluence page, space, or wiki section.
• User pastes a wiki URL or a numeric pageId.

Prefer MCP tools when the confluence server is connected.

This connector is read-only. Do not promise to create or edit pages.

---

Auth

CONFLUENCE_BASE_URL and CONFLUENCE_PAT live in ~/.cursor/mcp.json → mcpServers.confluence.env. Never commit. Never echo the token.

Server / Data Center only. Confluence Cloud is out of scope.

{
  "mcpServers": {
    "confluence": {
      "command": "node",
      "args": ["/absolute/path/to/confluence-cursor-connector/dist/stdio-entry.js"],
      "env": {
        "CONFLUENCE_BASE_URL": "https://wiki.example.com",
        "CONFLUENCE_PAT": "…"
      }
    }
  }
}

After code changes: npm run build, then Reload MCP.

---

Tools

Tool	Purpose
get-current-user	Auth smoke test
list-spaces	Spaces the PAT can see (limit 1–100)
search-pages	CQL search; returns ids/titles/urls, not body
get-page	Full page: page_id or space_key + exact title

How to resolve a page

1. User gave a URL with pageId= or /pages/123 → get-page with page_id.
2. User gave space key + title → get-page with both.
3. User gave a topic → search-pages with CQL, then get-page on the best id.
4. If several hits — list them, ask once.

CQL examples:

title~"onboarding" AND type=page
space = PRODUCT AND type = page
text ~ "лимиты" AND type = page

---

Reading bodyStorage

get-page returns Confluence storage HTML, not Markdown.

Before quoting or summarizing:

1. Strip tags and ac: macros.
2. Keep headings, lists, tables as structure.
3. If a macro hid the real content (excerpt, include, unexpanded drawio) — say so and point the user to the URL from the tool result.

Do not invent page text that was not in bodyStorage.

---

Agent workflow