# Building a Confluence MCP Server for Claude Code: From Setup to Skills

Tariqul Islam Ronnie — Tue, 16 Jun 2026 14:32:23 +0000

Automate your Confluence workflow directly from Claude Code — publish pages, sync content, and upload images without leaving your terminal.

GitHub: https://github.com/tariqulislam/confluence-mcp-server

What is an MCP Server?

MCP stands for Model Context Protocol — an open standard that lets AI assistants like Claude Code connect to external tools and services. An MCP server acts as a bridge: it exposes a set of tools that Claude Code can call, allowing it to interact with third-party systems (like Confluence) on your behalf.

In this guide you will learn:

How the Confluence MCP server is structured
How environment variables wire credentials into the server
How the server is registered with Claude Code
How Claude Code Skills wrap the tools into friendly slash commands
How to use every skill end-to-end

Project Structure

confluence-mcp-mock/
├── .env                     ← your credentials (never committed)
├── .env.example             ← template
├── .mcp.json                ← registers MCP servers with Claude Code
├── confluence-mcp/
│   ├── src/
│   │   ├── index.ts         ← entry point
│   │   ├── server.ts        ← tool registration
│   │   ├── config.ts        ← env validation (Zod)
│   │   ├── confluence/
│   │   │   ├── client.ts    ← Axios HTTP client
│   │   │   ├── api.ts       ← REST API calls
│   │   │   ├── parser.ts    ← XHTML to structured data
│   │   │   └── markdown-converter.ts
│   │   └── tools/           ← one file per MCP tool
│   │       ├── connection-check.ts
│   │       ├── list-pages.ts
│   │       ├── publish-page.ts
│   │       ├── read-save-page.ts
│   │       ├── upload-images.ts
│   │       └── sync-pages.ts
│   └── build/               ← compiled output
└── .claude/skills/          ← slash command definitions
    ├── confluence-conn-check/
    ├── confluence-page-list/
    ├── publish-to-confluence/
    ├── confluence-upload-images/
    └── confluence-sync-pages/

Step 1 — Configure Your .env File

The server reads all credentials from a single .env file at the project root. Copy the example template:

cp .env.example .env

Then fill in the four Confluence variables:

# .env  (never commit this file)
CONFLUENCE_URL=https://your-confluence-instance.com/confluence/
CONFLUENCE_USER=your.email@company.com
CONFLUENCE_PERSONAL_ACCESS_TOKEN=your_token_here
CONFLUENCE_SPACE_KEY=MYSPACE

The four variables you need:

CONFLUENCE_URL — Base URL of your Confluence instance. Must end with a trailing slash.
Example: https://confluence.company.com/confluence/

CONFLUENCE_USER — The email address tied to your Confluence account.

CONFLUENCE_PERSONAL_ACCESS_TOKEN — A Bearer token generated in Confluence settings.

CONFLUENCE_SPACE_KEY — The space key where new pages will be created. Example: ENGINEERING

How to generate a Personal Access Token

Log into Confluence and click your avatar → Profile
Navigate to Personal Access Tokens in the left sidebar
Click Create token, give it a name, set an expiry date
Copy the token immediately — it is shown only once

Step 2 — How the Server Loads Credentials

When the server starts, config.ts reads the .env file and validates every variable using Zod — a TypeScript schema library. If any variable is missing or malformed, the server exits with a clear error instead of failing silently later.

// confluence-mcp/src/config.ts  (simplified)
import dotenv from 'dotenv';
import { z } from 'zod';

dotenv.config({ path: join(__dirname, '..', '..', '.env') });

const envSchema = z.object({
  CONFLUENCE_URL: z.string().url(),
  CONFLUENCE_USER: z.string().email(),
  CONFLUENCE_PERSONAL_ACCESS_TOKEN: z.string().min(1),
  CONFLUENCE_SPACE_KEY: z.string().min(1),
});

export const config = envSchema.parse(process.env);

The validated config is imported by client.ts, which creates a single Axios instance used by every tool:

// confluence-mcp/src/confluence/client.ts  (simplified)
const client = axios.create({
  baseURL: config.CONFLUENCE_URL,
  headers: {
    Authorization: `Bearer ${config.CONFLUENCE_PERSONAL_ACCESS_TOKEN}`,
    Accept: 'application/json',
    'Content-Type': 'application/json',
  },
  timeout: 30000,
});

Credentials are set once here and never repeated anywhere else in the codebase.

Step 3 — Register the Server with Claude Code

Claude Code discovers MCP servers through the .mcp.json file in the project root:

{
  "mcpServers": {
    "confluence": {
      "command": "node",
      "args": ["./confluence-mcp/build/index.js"]
    }
  }
}

This tells Claude Code: "When this project is open, launch the node process at that path and communicate with it over stdio."

Step 4 — Build and Start

# Install dependencies
cd confluence-mcp && npm install

# Compile TypeScript to JavaScript
npm run build

# Restart Claude Code to load the new server

During development, use watch mode so the server recompiles on every save:

npm run watch

Step 5 — Understanding the Server Architecture

The server follows a clean three-layer pattern:

Claude Code
    │  (MCP protocol over stdio)
    ▼
server.ts         ← registers tools, routes calls
    │
    ▼
tools/*.ts        ← one file per tool, handles input/output
    │
    ▼
confluence/api.ts ← REST API calls (Confluence v1 endpoints)

How a tool is registered

server.ts declares every tool's name, description, and input schema in the ListToolsRequestSchema handler. Claude Code reads this to understand what tools exist. Calls are then routed in the CallToolRequestSchema handler:

// server.ts  (simplified)
server.setRequestHandler(ListToolsRequestSchema, async () => ({
  tools: [
    {
      name: 'confluence_connection_check',
      description: 'Verify connection to Confluence and check authentication.',
      inputSchema: { type: 'object', properties: {}, required: [] },
    },
    {
      name: 'confluence_list_pages',
      description: 'List all child pages under a parent page.',
      inputSchema: {
        type: 'object',
        properties: {
          parent_page_id: { type: 'string', description: 'ID of the parent page' },
        },
        required: ['parent_page_id'],
      },
    },
  ],
}));

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  switch (request.params.name) {
    case 'confluence_connection_check':
      return { content: [{ type: 'text', text: await handleConnectionCheck() }] };
    case 'confluence_list_pages':
      return { content: [{ type: 'text', text: await handleListPages(args) }] };
  }
});

Available MCP Tools

confluence_connection_check — Verifies auth and returns server/user info

confluence_list_pages — Lists all child pages under a parent

confluence_read_save_page — Downloads a page (sections, tables, images, comments) to local JSON

confluence_publish_page — Converts markdown to Confluence XHTML and creates/updates a page

confluence_upload_images — Bulk-uploads images from a local directory as page attachments

confluence_sync_pages — Syncs all child pages to local markdown with change tracking

Step 6 — Creating Skills (Slash Commands)

Skills are the user-friendly layer on top of raw MCP tools. A skill is a directory under .claude/skills/ containing a SKILL.md file.

.claude/skills/confluence-conn-check/
└── SKILL.md

SKILL.md format

---
name: confluence-conn-check
description: Verify Confluence server connection and authentication.
argument-hint: (no arguments required)
---

# Confluence Connection Check

Test the connection to Confluence and verify your credentials.

## Usage

/confluence-conn-check

## MCP Tool

confluence_connection_check

The three frontmatter fields that matter:

name — becomes the slash command (e.g. /confluence-conn-check)

description — shown in the / autocomplete menu and used for auto-discovery by Claude

argument-hint — hint shown in autocomplete, e.g. <parent_page_id> or (no arguments required)

When you type /confluence-conn-check in Claude Code, it reads SKILL.md, knows which MCP tool to call, and formats the response back to you.

Step 7 — Using Every Skill

/confluence-conn-check

Verify your credentials before doing anything else.

/confluence-conn-check

Expected response:

{
  "status": "success",
  "user": {
    "displayName": "Jane Doe",
    "email": "jane.doe@company.com"
  },
  "server": {
    "version": "8.5.0",
    "baseUrl": "https://confluence.company.com/confluence"
  }
}

/confluence-page-list <parent_page_id>

List all child pages under a parent so you can find page IDs.

/confluence-page-list 123456789

Response:

{
  "status": "success",
  "message": "Found 3 child page(s)",
  "pages": [
    { "id": "987654321", "title": "Getting Started" },
    { "id": "987654322", "title": "API Reference" }
  ]
}

/confluence-page-read-save-content <page_id>

Download a full page — sections, tables, images, and comments — and save it locally as structured JSON.

/confluence-page-read-save-content 987654321

Saves to:

confluence-docs/con_<parentId>_<pageId>/
├── sections_and_sub_sections_contents.json
├── comments_contents.json
└── images/
    ├── screenshot.png
    └── diagram.jpg

/publish-to-confluence <parent_id> <file.md> [page_id]

Convert a markdown file to Confluence storage format and publish it. Images referenced in the markdown are automatically uploaded as attachments.

Create a new page:

/publish-to-confluence 123456789 docs/architecture.md

Update an existing page:

/publish-to-confluence 123456789 docs/architecture.md 987654321

The markdown converter handles headings, bold/italic, fenced code blocks with syntax highlighting, ordered and unordered lists (nested), pipe tables, local and external images, blockquotes, and horizontal rules.

/confluence-upload-images <page_id> <dir> [batch_size] [delay_ms]

Bulk-upload images from a local directory with configurable rate-limit control.

# Default (3 parallel, 2 s delay between batches)
/confluence-upload-images 987654321 ./screenshots

# Conservative — slow instance or large files
/confluence-upload-images 987654321 ./screenshots 1 3000

Rate-limit tuning guide:

batch_size=1, delay_ms=3000 — safest, use on aggressive instances
batch_size=3, delay_ms=2000 — default, works for most instances
batch_size=5, delay_ms=1000 — faster, for lightly loaded on-premise servers

/confluence-sync-pages <parent_page_id>

Download all child pages under a parent as local markdown files. On repeat runs, only pages that actually changed (version bump or new attachment) are re-downloaded.

/confluence-sync-pages 6656101928

Output structure:

confluence_doc_md/
└── parent_6656101928/
    ├── manifest.json      ← change-tracking index
    ├── page_987654321/
    │   ├── page.md
    │   ├── comments.json
    │   └── images/
    └── page_987654322/
        └── page.md

Each page.md starts with a metadata header:

# Page Title

> Page ID: 987654321
> Version: 5
> Last modified: 2026-05-20T10:30:00.000Z

[page body]

---

## Comments

### Comment by jane.doe (2026-05-19T09:00:00.000Z)

Comment text here.

Change detection logic:

First sync (no manifest) → download everything
Page version number increased → re-download body, comments, attachments
Attachment has newer timestamp → re-download attachments
Nothing changed → skip the page entirely

End-to-End Workflow Example

# 1. Verify credentials
/confluence-conn-check

# 2. Find the parent page ID
/confluence-page-list 100000001

# 3. Write docs locally in your editor
# (create docs/my-feature.md)

# 4. Publish
/publish-to-confluence 100000002 docs/my-feature.md

# 5. Update after edits
/publish-to-confluence 100000002 docs/my-feature.md 987654321

# 6. Back up all pages under a parent
/confluence-sync-pages 100000001

Adding a New Tool

To extend the server with a custom tool, follow four steps.

1. Create the tool file at confluence-mcp/src/tools/your-tool.ts:

import { z } from 'zod';

export const yourToolSchema = z.object({
  page_id: z.string().describe('The Confluence page ID'),
});

export async function handleYourTool(input: z.infer<typeof yourToolSchema>) {
  return JSON.stringify({ status: 'success', data: {} }, null, 2);
}

2. Register in server.ts — add to the tool list and the switch statement:

import { handleYourTool } from './tools/your-tool.js';

// in ListToolsRequestSchema handler:
{ name: 'confluence_your_tool', description: '...', inputSchema: { ... } }

// in CallToolRequestSchema handler:
case 'confluence_your_tool':
  return { content: [{ type: 'text', text: await handleYourTool(args) }] };

3. Create the skill at .claude/skills/your-tool/SKILL.md following the format in Step 6.

4. Rebuild and restart:

cd confluence-mcp && npm run build
# Restart Claude Code

Troubleshooting

MCP server not loading
Run npm run build inside confluence-mcp/ and check for TypeScript errors. Verify .mcp.json points to ./confluence-mcp/build/index.js. Fully restart Claude Code — the server process is launched at startup.

401 Authentication Error
Regenerate your Personal Access Token in Confluence settings. Confirm CONFLUENCE_USER is the email associated with the token.

URL errors
CONFLUENCE_URL must end with a trailing slash. If behind a VPN, connect before running.

Publishing formatting issues
Tables need a header separator row (|---|---|). Avoid emoji and raw angle bracket characters (<, >) directly inside table cells — Confluence's XHTML parser rejects them.

Rate limit exceeded during image upload
Slow down: /confluence-upload-images <page_id> <dir> 1 4000

Summary

The full stack in one view:

.env — credentials, never committed
config.ts — Zod validates env at startup
client.ts — single Axios instance with Bearer auth
api.ts — Confluence REST v1 endpoint calls
tools/*.ts — one handler per MCP tool
server.ts — declares tools to Claude Code via MCP protocol
.mcp.json — tells Claude Code where to find the server
.claude/skills/*/SKILL.md — slash commands that wrap MCP tools

With this architecture you can publish markdown documentation, sync entire Confluence spaces, and upload images — all from a single Claude Code session.

Source code: github.com/tariqulislam/confluence-mcp-server

DEV Community: Tariqul Islam Ronnie