Enhanced Markdown is now an MCP server for AI agents
Markdown won because it is the format AI models write best: it is plain text, it diffs cleanly, and every developer already has it in a repo somewhere. The wiki is where that content is supposed to end up — architecture docs, runbooks, RFCs, meeting notes — and "import markdown to Confluence" is a search people type often enough that it stings. The distance between "a markdown file" and "a Confluence page" has always been a copy-paste with formatting losses at the end.
The MCP era should close that loop. This week it did for diagrams — Mermaid Plus became an MCP server. Now the document layer follows.
What we shipped
Enhanced Markdown for Confluence now ships a rovo:mcp module — Atlassian's bridge that turns Forge app actions into MCP tools. The same four actions that power the built-in Rovo agent are exposed as an MCP server:
| Tool | What it does |
|---|---|
create-markdown | Adds a markdown block to a page — or creates a new page/blogpost from it |
get-markdown | Reads back the markdown source and the page version |
update-markdown | Replaces a markdown block with a version-checked write |
convert-markdown-blocks | Turns markdown code blocks already on the page into rendered blocks, in place |
These edit the real content on the real page. Everything the macro renders in the editor works through the tools: tables, fenced code with syntax highlighting, KaTeX math, Mermaid fences. Version locking prevents lost updates; an open editor draft is synced instead of clobbered; every call runs with the invoking user's own Confluence permissions.
Using it in Claude Code, end to end
This is the part you came for. One-time setup:
1. Admin: expose the tools. External access is off by default. In admin.atlassian.com, open your organization → Apps → Sites → your site → Connected apps → Enhanced Markdown for Confluence, and switch on the four markdown tools. Once per site.
2. Register the server in Claude Code:
claude mcp add --scope user --transport http enhanced-markdown \
https://mcp.atlassian.com/forge/6c4adb40-91c3-41db-8804-db25b2998069
claude mcp list
# enhanced-markdown: https://mcp.atlassian.com/forge/6c4adb40-… - ! Needs authentication
3. Authenticate. Start claude, run /mcp, select enhanced-markdown → Authenticate. A browser opens the Atlassian OAuth 2.1 consent screen — sign in, pick your site, approve. claude mcp list flips to ✔ Connected.
4. Work your wiki. Four prompts, one per tool — paste-ready:
Create a page titled "API Style Guide" in our DEV space from this markdown: …(paste)…
Read the markdown on https://your-site.atlassian.net/wiki/spaces/TEAM/pages/123456789/Architecture and tighten the wording
Update the markdown on page 123456789 — replace the Draft warning with the final release notes below: …(paste)…
Convert the markdown code blocks on https://your-site.atlassian.net/wiki/spaces/TEAM/pages/987654321/Imported-Notes into rendered blocks
Two things to know: over MCP there is no open page, so point the client at the page URL or its numeric ID (the number in …/pages/123456789/Title) or a space key — the tools' errors say exactly this rather than guessing a target. And until the OAuth step is done, non-interactive Claude runs skip the server silently; check claude mcp list first.
Cursor users: add the same URL as a remote HTTP MCP server — the OAuth flow is identical. The full walkthrough, including the admin setup and the Rovo Studio path, lives in the user guide.
Multi-block pages: the tools refuse to guess
A wiki page can hold several markdown macros — a spec and an appendix, a report and its data notes. So get-markdown and update-markdown stop and ask which block you mean whenever a page has more than one, numbered from the top. Your answer becomes the index, and the write lands exactly there. The one failure a wiki-writing agent must never have is quietly editing the wrong block; we traded a round-trip for that guarantee.
The same philosophy runs through the whole surface: version mismatches are reported, never overwritten; an open editor draft is written into and republished, never clobbered by a three-way merge; and if a request could mean two different pages, the tool asks instead of picking.
Why an app, not a standalone MCP server
A standalone Confluence MCP server would need you to hand a third party an API token with write access to your wiki. Ours rides on the Forge app you already installed: Atlassian hosts it, access goes through Atlassian's own OAuth and permission layer, and admin controls exposure per site. The same argument we made for diagrams applies to documents — the full version is in the Mermaid Plus post.
With both apps exposing MCP, one agent can hold the whole page: "update the architecture doc, then regenerate its sequence diagram" is now two tool calls in one conversation.
What's next
Search is the obvious next tool — finding the right page without leaving the agent. If you're building agents on top of your wiki, tell us what's missing: the Marketplace listing takes feedback, and so does this blog.