Confluence Mini

Lightweight MCP server that crawls Confluence subtrees and serves cached Markdown. Works with Claude Code and nanobot (Slack) via stdio.

Claude Code config

{
  "mcpServers": {
    "confluence-docs": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/bonkey/confluence-mini-mcp", "confluence-mini-mcp"],
      "env": {
        "CONFLUENCE_BASE_URL": "https://company.atlassian.net/wiki",
        "CONFLUENCE_EMAIL": "[email protected]",
        "CONFLUENCE_API_TOKEN": "...",
        "CONFLUENCE_ROOT_PAGE_IDS": "123456,789012"
      }
    }
  }
}

Pin to a tag: "--from", "git+https://github.com/bonkey/[email protected]"

Nanobot config

[mcp.confluence]
command = "uvx"
args = ["--from", "git+https://github.com/bonkey/confluence-mini-mcp", "confluence-mini-mcp"]
env = { CONFLUENCE_BASE_URL = "https://company.atlassian.net/wiki", CONFLUENCE_EMAIL = "[email protected]", CONFLUENCE_API_TOKEN = "...", CONFLUENCE_ROOT_PAGE_IDS = "123456,789012" }

Tools

Crawling

The crawler walks a graph, not just a tree:

• Child pages — standard parent → children traversal
• Linked Confluence pages — any internal link in page content is followed (enables a single "hub page" linking to multiple subtrees)
• External HTTP links — fetched as Markdown (1 level deep, follows redirects only)

In-page directives

Add a [confluence-mcp] block anywhere on a Confluence page to control crawl behaviour for that subtree:

[confluence-mcp]
max_depth=3
follow_links=false
follow_external=true

The directive block is stripped from the Markdown output — agents never see it.

Configuration

All settings via env vars or confluence-mini-mcp.toml:

Development

just dev     # run in dry-run mode
just check   # format + test
just release # tag + push

All tasks in the Justfile. Requires uv and just.