ultra-confluence-mcp

A Model Context Protocol (MCP) server for Confluence Cloud, designed around one question: how much of the agent's context window does a Confluence call actually need to consume?

Most Confluence MCP servers pass the Confluence REST API through more or less verbatim. That's fine for occasional use, but a single getPage on a long doc can dump 40 KB of ADF JSON into the conversation, and a 25-page list response runs over 500 KB. Agents pay that cost on every call, and it crowds out the actual work.

This server takes a different stance.

Why use this over other Confluence/Atlassian MCPs

1. Per-call responses are trimmed by ~20× on average

A built-in projection layer drops _links, _expandable, formatter noise, and other fields agents never read, and converts ADF/storage bodies to markdown on the way out. Concrete numbers from the benchmark in docs/BENCHMARK.md:

2. Large bodies offload to disk, not into context

When a page body would exceed the inline limit, the trim layer writes the raw API response to a temp file and returns a tiny bodyPath reference (~500 bytes) instead. Agents that need the content call confluence_render_body to read from disk — and can pass outputPath to write the rendered markdown straight to a file, so the body bytes never enter the conversation at all. Restoring five 10 KB pages costs zero context bytes with this path; an MCP that inlines bodies pays ~50 KB.

3. The per-conversation tool surface is small and tunable

The MCP tool-list response itself costs context — every conversation pays for it up front, before any work happens. With CONFLUENCE_ENABLED_CATEGORIES you can scope the surface to what an agent actually needs:

By comparison, many combined Atlassian MCP servers expose over 70 tools across Jira and Confluence — fine if you want both products, but a lot of schema for an agent that just needs to read and write Confluence pages. If you also need Jira or Bitbucket, the companion servers ultra-jira-mcp and ultra-bitbucket-mcp apply the same trimming philosophy — wire up only the ones you need instead of paying for one monolithic Atlassian surface.

4. There's a CLI for agents that prefer shelling out

confluence-cli is a standalone binary that calls Confluence directly — no MCP host required. Same trimmed output shape, plus a ref: /path line pointing at the full untrimmed response on disk. Per-conversation overhead is a single 2.3 KB SKILL.md loaded on demand by the Claude Code harness, versus ~50 KB of tool schemas. The two paths share all the trim logic, so you can mix or switch without behavior drift.

5. No Docker, no Python, no runtime to install

It's a Node package. If you already have Node (you probably do — Claude Code ships with it), npx -y https://github.com/scottlepp/ultra-confluence-mcp is the whole install. No container to pull, no Python virtualenv to manage, no uv/pipx/poetry to learn first, no docker run line with seven -e flags in your MCP config. Drop the npx command into Claude Desktop / Claude Code / Cursor and you're done.

6. Self-healing — humans aren't the bottleneck (without being a security liability)

The repo is maintained by bots, so updates don't stall waiting on a human reviewer — but every code-modifying step that touches an LLM is gated so a malicious issue or PR can't turn the automation into an exfiltration channel:

• Dependencies stay current. Dependabot opens grouped PRs on a weekly cadence. Patch and minor bumps auto-merge once CI is green. Major bumps fail the workflow and require a human reviewer — the previous behavior of auto-running an LLM "migration agent" on PR code is disabled, because executing PR-supplied code in a privileged context is an RCE pattern.

• Bug reports get triaged automatically — when a maintainer approves. The scheduled bug-fix agent only processes issues bearing the auto-fix-approved label, which is restricted to maintainers. Issue bodies are treated as untrusted text (boundary-tagged in the prompt, truncated, scanned for shell metacharacters and secret shapes). The agent opens PRs but never auto-merges them; a human still reviews and clicks merge.

• PRs get a first-pass review without waiting on a human. The PR-review agent runs against the diff as data — it never checks out PR code into the workspace where the privileged token bag lives. Agent code is always loaded from main.

• The agent's tools are sandboxed. File reads and writes go through a shared path policy that rejects absolute paths, .. traversal, symlinks pointing outside the working dir, and a denylist that covers .env*, .github/, scripts/agents/ itself, lockfiles, key files, and similar high-blast-radius paths. Git and npm commands use execFile with argv arrays — no shell interpolation, so model-supplied branch names and commit messages cannot inject $() or backticks. See scripts/agents/src/validation/ for the policy module and its test coverage.

This matters for a context-efficiency tool specifically: the value proposition decays fast if the trim layer falls behind a Confluence API change or a CVE in a dependency. Self-healing keeps the surface fresh without a maintainer in the loop — and without giving the loop a way to turn into a backdoor.

What this server is NOT

• Not multi-product. Jira lives in ultra-jira-mcp and Bitbucket in ultra-bitbucket-mcp; this server is Confluence Cloud only, by design to keep the tools light.

• Not Server/Data Center. Cloud REST API v2 only.

• Not for human-readable conversations. The trimming is aggressive because agents read JSON, not docs — if you want pretty browser-style output, use the Atlassian web UI.

If you need Jira + Confluence + Bitbucket + Server/DC + OAuth in one server, there are MCP servers for that, but you'll pay the price. If your agents are blowing through context windows on Confluence reads, use this — and pair it with ultra-jira-mcp or ultra-bitbucket-mcp when you need those too.

Installation

Run directly from GitHub:

npx -y https://github.com/scottlepp/ultra-confluence-mcp

Or clone and build locally:

git clone https://github.com/scottlepp/ultra-confluence-mcp.git
cd ultra-confluence-mcp
npm install
npm run build
node build/index.js

Standalone CLI (no MCP server)

A confluence-cli binary ships alongside the MCP server. It reads the same env vars (CONFLUENCE_HOST, CONFLUENCE_EMAIL, CONFLUENCE_API_TOKEN), builds a Confluence client in-process, and calls Confluence directly — no MCP host required.

export CONFLUENCE_HOST=https://yourcompany.atlassian.net
export CONFLUENCE_EMAIL=you@example.com
export CONFLUENCE_API_TOKEN=...

npx -y -p github:scottlepp/ultra-confluence-mcp confluence-cli confluence_get_page --pageId=12345

Discovery:

confluence-cli --help                       # list every tool
confluence-cli <tool> --help                # show flags for one tool

Flag forms: --key=value, --key value, --key=@/path/to/file (read from file), --key=- (read from stdin), repeated --key=a --key=b to build an array (or comma-separated --key=a,b).

On success the CLI prints a trimmed JSON summary to stdout, then a final ref: /tmp/.../...json line pointing at the full untrimmed response on disk — cat it when the summary leaves out detail you need. Pass --full=true to skip trimming and dump the raw response inline instead.

Claude Code skill

To make the CLI discoverable to Claude Code agents in standalone sessions, install the bundled skill:

npx -y -p github:scottlepp/ultra-confluence-mcp confluence-cli install-skill

This writes ~/.claude/skills/confluence/SKILL.md. Use --force to overwrite, or --print to dump the skill content to stdout without writing.

MCP vs CLI — which should I use?

There's a popular claim that CLIs are categorically more context-efficient than MCP. The benchmark in docs/BENCHMARK.md disagrees: with aggressive tool filtering (CONFLUENCE_ENABLED_CATEGORIES), the MCP path's per-conversation overhead is within a few KB of the CLI's SKILL.md. Per-call output is identical in both paths (same trimmed shape from the same applyTrim projection). Without filtering, the CLI's footprint is ~22× smaller than the full MCP tool list — but anyone who's filtering is already in the same ballpark.

The CLI's real edge is ergonomic: scriptable, pipeable, doesn't require an MCP host, runs from any shell. Use whichever fits your workflow; on token cost it's mostly a wash for reasonable filter configs.

Configuration

Set the following environment variables:

Tool Filtering

You can limit which tools are exposed to the AI model using environment variables:

Enable only specific categories:

CONFLUENCE_ENABLED_CATEGORIES=page,space,search

Disable specific tools (e.g., destructive operations):

CONFLUENCE_DISABLED_TOOLS=confluence_delete_page,confluence_delete_space,confluence_delete_blog_post

Available categories: page, space, blogPost, comment, attachment, label, search, user, version, contentProperty, ancestor, descendant, server

Claude Desktop Setup

Add to your Claude Desktop configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "confluence": {
      "command": "npx",
      "args": ["-y", "https://github.com/scottlepp/ultra-confluence-mcp"],
      "env": {
        "CONFLUENCE_HOST": "https://yourcompany.atlassian.net",
        "CONFLUENCE_EMAIL": "your-email@example.com",
        "CONFLUENCE_API_TOKEN": "your-api-token"
      }
    }
  }
}

With tool filtering (recommended for limited access):

{
  "mcpServers": {
    "confluence": {
      "command": "npx",
      "args": ["-y", "https://github.com/scottlepp/ultra-confluence-mcp"],
      "env": {
        "CONFLUENCE_HOST": "https://yourcompany.atlassian.net",
        "CONFLUENCE_EMAIL": "your-email@example.com",
        "CONFLUENCE_API_TOKEN": "your-api-token",
        "CONFLUENCE_ENABLED_CATEGORIES": "page,space,search,comment",
        "CONFLUENCE_DISABLED_TOOLS": "confluence_delete_page,confluence_delete_space"
      }
    }
  }
}

Available Tools

Pages

• confluence_get_pages - Get all pages with optional filters

• confluence_get_page - Get a specific page by ID

• confluence_create_page - Create a new page

• confluence_update_page - Update an existing page

• confluence_delete_page - Delete a page

• confluence_get_pages_in_space - Get all pages in a space

• confluence_get_pages_for_label - Get pages with a specific label

• confluence_render_body - Convert a cached page body (ADF/storage on disk) to markdown. Pass outputPath to write straight to disk so the rendered body never inflates the agent's context. See Reading large bodies.

Spaces

• confluence_get_spaces - List all accessible spaces

• confluence_get_space - Get space details

• confluence_create_space - Create a new space

• confluence_update_space - Update space

• confluence_delete_space - Delete space

Blog Posts

• confluence_get_blog_posts - Get all blog posts

• confluence_get_blog_post - Get blog post details

• confluence_create_blog_post - Create a new blog post

• confluence_update_blog_post - Update blog post

• confluence_delete_blog_post - Delete blog post

• confluence_get_blog_posts_in_space - Get blog posts in a space

Comments

• confluence_get_page_footer_comments - Get footer comments on a page

• confluence_get_page_inline_comments - Get inline comments on a page

• confluence_get_blog_post_footer_comments - Get footer comments on a blog post

• confluence_get_footer_comment - Get a specific comment

• confluence_create_page_footer_comment - Add comment to page

• confluence_create_blog_post_footer_comment - Add comment to blog post

• confluence_update_footer_comment - Update comment

• confluence_delete_footer_comment - Delete comment

Attachments

• confluence_get_page_attachments - Get attachments on a page

• confluence_get_blog_post_attachments - Get attachments on a blog post

• confluence_get_attachment - Get attachment details

• confluence_delete_attachment - Delete attachment

Labels

• confluence_get_page_labels - Get labels on a page

• confluence_add_page_label - Add label to page

• confluence_remove_page_label - Remove label from page

• confluence_get_blog_post_labels - Get labels on a blog post

• confluence_add_blog_post_label - Add label to blog post

• confluence_remove_blog_post_label - Remove label from blog post

• confluence_get_space_labels - Get labels on a space

• confluence_add_space_label - Add label to space

• confluence_remove_space_label - Remove label from space

Search

• confluence_cql_search - Advanced search using CQL (Confluence Query Language) for pages, blog posts, attachments, comments

• confluence_search_content - Simple text search for pages and blog posts

• confluence_search_generic_content - Search for databases, whiteboards, folders, or embeds (NOT for pages/blog posts)

Users

• confluence_get_current_user - Get authenticated user

• confluence_get_user - Get user by account ID

• confluence_get_users - Get all users

Versions

• confluence_get_page_versions - Get page version history

• confluence_get_page_version - Get specific page version

• confluence_get_blog_post_versions - Get blog post version history

• confluence_get_blog_post_version - Get specific blog post version

Content Properties

• confluence_get_page_properties - Get content properties for a page

• confluence_get_page_property - Get a specific property

• confluence_create_page_property - Create a property

• confluence_update_page_property - Update a property

• confluence_delete_page_property - Delete a property

Ancestors & Descendants

• confluence_get_page_ancestors - Get parent pages

• confluence_get_page_descendants - Get all descendant pages

• confluence_get_page_children - Get direct children pages

Server

• confluence_get_server_info - Get Confluence server info

Available Resources

• confluence://spaces - List of all accessible spaces

• confluence://myself - Current user info

• confluence://space/{id} - Space details

• confluence://page/{id} - Page details

• confluence://blogpost/{id} - Blog post details

Reading large bodies without context bloat

When a single-page read returns a body too large to inline, the trim layer offloads the raw API response to disk and surfaces a bodyPath field on the response. Two ways to use it:

Inline rendering (default). Call confluence_render_body with just bodyPath to get the converted markdown back in the response under bodyMarkdown:

{
  "bodyPath": "/var/folders/.../ultra-confluence-mcp/pages/12345-v3.json"
}

Direct-to-disk rendering (recommended when the agent already knows where the file should land — e.g. pulling a doc back to a working tree). Add outputPath and the rendered output is written there; the response carries only { representation, sourceLength, outputPath, bytesWritten } — the body bytes never traverse the agent's context:

{
  "bodyPath": "/var/folders/.../ultra-confluence-mcp/pages/12345-v3.json",
  "outputPath": "/abs/path/to/page.md"
}

This is dramatically more context-efficient when restoring multiple pages. Five docs of ~10 KB each cost ~50 KB through inline rendering and effectively 0 through outputPath. Parent directories are created if missing; existing files are overwritten.

Both forms work with format: "raw" if you want the original ADF JSON / storage XHTML rather than markdown.

Body Formats

Confluence supports multiple body formats:

• storage - XHTML-based storage format (default)

• atlas_doc_format - Atlassian Document Format (ADF) - supports Forge app extensions

• view - Rendered HTML (read-only)

Storage Format (default)

<p>This is a paragraph.</p>
<h1>This is a heading</h1>
<ul>
  <li>List item 1</li>
  <li>List item 2</li>
</ul>

ADF Format

Use bodyFormat: "atlas_doc_format" when creating/updating pages to insert Forge app macros (like Mermaid diagrams). The body should be a JSON-stringified ADF document:

{
  "version": 1,
  "type": "doc",
  "content": [
    {
      "type": "paragraph",
      "content": [{ "type": "text", "text": "Hello world" }]
    }
  ]
}

See Inserting Forge App Macros for advanced usage.

Inserting Forge App Macros (e.g., Mermaid Diagrams)

You can programmatically insert Forge app macros like Mermaid diagrams using the ADF (Atlassian Document Format) body format.

Mermaid Diagram Example

To insert a Mermaid diagram, use bodyFormat: "atlas_doc_format" with both a code block and an extension node:

{
  "spaceId": "123456",
  "title": "Page with Mermaid Diagram",
  "bodyFormat": "atlas_doc_format",
  "body": "{\"version\":1,\"type\":\"doc\",\"content\":[{\"type\":\"codeBlock\",\"attrs\":{\"language\":\"mermaid\"},\"content\":[{\"type\":\"text\",\"text\":\"sequenceDiagram\\n    Alice->>Bob: Hello\\n    Bob-->>Alice: Hi!\"}]},{\"type\":\"extension\",\"attrs\":{\"extensionKey\":\"23392b90-4271-4239-98ca-a3e96c663cbb/63d4d207-ac2f-4273-865c-0240d37f044a/static/mermaid-diagram\",\"extensionType\":\"com.atlassian.ecosystem\",\"parameters\":{\"localId\":\"mermaid-1\"},\"localId\":\"mermaid-1\"}}]}"
}

ADF Structure for Mermaid

The ADF document structure (before JSON stringification):

{
  "version": 1,
  "type": "doc",
  "content": [
    {
      "type": "codeBlock",
      "attrs": { "language": "mermaid" },
      "content": [
        { "type": "text", "text": "sequenceDiagram\n    Alice->>Bob: Hello" }
      ]
    },
    {
      "type": "extension",
      "attrs": {
        "extensionKey": "23392b90-4271-4239-98ca-a3e96c663cbb/63d4d207-ac2f-4273-865c-0240d37f044a/static/mermaid-diagram",
        "extensionType": "com.atlassian.ecosystem",
        "parameters": { "localId": "mermaid-1" },
        "localId": "mermaid-1"
      }
    }
  ]
}

Key points:

• The extensionKey is for the Mermaid diagrams viewer app by Atlassian Labs (must be installed on your Confluence instance)

• The localId must be in both parameters.localId AND attrs.localId

• The Mermaid macro auto-detects code blocks by position (1st extension → 1st code block, 2nd → 2nd, etc.)

• Use "parameters": { "index": N } to explicitly select which code block (0-based index)

Finding Extension Keys for Other Forge Apps

To find the extension key for other Forge apps:

1. Manually insert the macro on a Confluence page

2. Fetch the page with bodyFormat: "atlas_doc_format"

3. Look for the extensionKey in the extension node

CQL Search Examples

The confluence_cql_search tool uses Confluence Query Language (CQL):

# Search by content type
type=page AND space=DEV

# Full-text search
text~"search term"

# Search by creator
creator=currentUser()

# Search by date
created>=now("-7d")

# Combined search
type=page AND space=DEV AND text~"api" AND created>=now("-30d")

Development

# Install dependencies
npm install

# Build
npm run build

# Run tests
npm test

# Run with inspector
npm run inspector

Self-Healing Agents

This repository includes AI agents for automated maintenance in scripts/agents/:

• BugFixAgent: Reads bug issues, validates them, implements fixes, and creates PRs

• PRReviewAgent: Reviews pull requests, identifies issues, and suggests improvements

Running Agents

cd scripts/agents
npm install

# Run bug fix agent
GITHUB_TOKEN=xxx ISSUE_NUMBER=123 npm run bug-fix

# Run PR review agent
GITHUB_TOKEN=xxx PR_NUMBER=123 PR_DIFF_FILE=diff.txt npm run pr-review

Supported AI Providers

The agents support multiple AI providers with automatic fallback:

• Google AI (Gemini)

• OpenAI (GPT-4)

• Anthropic (Claude)

• Groq

• Mistral

• Perplexity

• DeepSeek

• OpenRouter

Set at least one API key:

GOOGLE_API_KEY=xxx
OPENAI_API_KEY=xxx
ANTHROPIC_API_KEY=xxx
# etc.

License

MIT