Note: This is still under development and is a preview. Currently testing on Qwen3 Coder Next REAP 40B A3B from here https://huggingface.co/lovedheart/Qwen3-Coder-Next-REAP-40B-A3B-GGUF/resolve/main/Qwen3-Coder-Next-REAP-40B-A3B-Q4_K_XL.gguf

Qwen3 MCP Server

A complete MCP server that gives LM Studio's Qwen3 (or any local LLM) full coding agent capabilities including 80+ tools for file operations, command execution, git, web search, memory, planning, and a full skills system.

Features

• Two Server Modes:

  • HTTP Mode - Browser-based chat at http://localhost:3847/chat.html
  • MCP Mode - Direct integration with LM Studio's MCP interface

• 80+ Tools including:

  • File operations (read, write, edit, search, glob)
  • Command execution (shell, background processes)
  • Git operations (status, diff, commit, branch, push, pull)
  • Web tools (search, image search, URL fetch)
  • Memory & planning (notes, scratchpads, task lists, plans)
  • Media (read images, PDFs, screenshots)
  • Jupyter notebook support
  • ComfyUI integration (32 workflow tools)
  • GitHub Blog (Jekyll blog creation, themes, deployment)

• Tool Name Aliases - Server-side routing handles common model hallucinations (e.g., edit -> edit_file, bash -> execute_command)

• Skills System - 16 pre-installed skills + install more from awesome-agent-skills

Quick Start

HTTP Mode (Browser Chat)

# Start the server
start-chat.bat

# Open in browser
http://localhost:3847/chat.html

MCP Mode (LM Studio Direct)

Add to your LM Studio MCP config (~/.lmstudio/mcp-servers.json):

{
  "mcpServers": {
    "qwen3-mcp": {
      "command": "node",
      "args": ["<YOUR_PATH>/src/index.js"],
      "cwd": "<YOUR_PATH>"
    }
  }
}

Replace <YOUR_PATH> with the actual path where you installed the server.

Then restart LM Studio.

Installation

Prerequisites

• Node.js 18+
• LM Studio with a model loaded (Qwen3 recommended)

Quick Setup (Windows)

cd <YOUR_PATH>
setup.bat

This will:

1. Install npm dependencies
2. Create required directories
3. Configure LM Studio MCP automatically

Manual Setup

# Install dependencies
npm install

# Setup MCP config for LM Studio
npm run setup

npm Scripts

Command	Description
npm run setup	Configure LM Studio MCP
npm run start	Start HTTP server (port 3847)
npm run mcp	Run MCP server directly (stdio)
npm run dev	Start with auto-reload

Project Structure

qwen3-mcp/
├── src/
│   ├── index.js           # MCP server (stdio mode)
│   └── tools/             # Tool implementations
│       ├── filesystem.js  # File operations
│       ├── edit.js        # File editing
│       ├── bash.js        # Command execution
│       ├── git.js         # Git operations
│       ├── search.js      # Glob/grep search
│       ├── web.js         # Web search, fetch
│       ├── memory.js      # Notes, scratchpads
│       ├── planning.js    # Plans, task tracking
│       ├── tasks.js       # Todo lists
│       ├── thinking.js    # Reasoning tools
│       ├── context.js     # Conversation context
│       ├── interaction.js # User prompts
│       ├── media.js       # Images, PDFs
│       ├── notebook.js    # Jupyter support
│       ├── comfyui.js     # ComfyUI workflows
│       ├── github-blog.js # Jekyll blog tools
│       └── skills.js      # Skills system
├── frontend/
│   ├── server.js          # HTTP server (port 3847)
│   ├── chat.html          # Browser chat interface
│   └── index.html         # Image viewer interface
├── skills/                # Installed skills (16)
│   ├── chrome-extension/  # Chrome extension development (MV3)
│   ├── code-review/       # Code review methodology
│   ├── comfyui-nodes/     # ComfyUI custom node development
│   ├── comfyui-workflow/  # ComfyUI workflow creation
│   ├── differential-review/ # Security-focused diff review
│   ├── docx/              # Word document creation
│   ├── frontend-design/   # Frontend UI/UX
│   ├── github-blog/       # Jekyll blog for GitHub Pages
│   ├── mcp-builder/       # Build MCP servers
│   ├── modern-python/     # Python tooling (uv, ruff)
│   ├── react-best-practices/ # React patterns (Vercel)
│   ├── shadcn-ui/         # Modern component library
│   ├── static-analysis/   # CodeQL, Semgrep, SARIF
│   ├── testing-handbook-skills/ # Fuzzers, sanitizers
│   ├── web-artifacts-builder/ # HTML/React prototypes
│   └── web-design-guidelines/ # UI/UX fundamentals
├── start-chat.bat         # Start HTTP server
├── stop-chat.bat          # Stop server
├── restart-chat.bat       # Restart server
└── install-skill.bat      # Install skills from GitHub

Available Tools (80+)

File Operations

Tool	Description
read_file	Read file contents with line numbers (params: file_path, offset, limit)
write_file	Write/create files (params: file_path, content)
edit_file	Find and replace in files (params: file_path, old_string, new_string)
list_directory	List directory contents (params: path)
create_directory	Create directories (params: path)
delete_file	Delete files (params: path)
move_file	Move/rename files (params: source, destination)
copy_file	Copy files (params: source, destination)
file_info	Get file metadata (params: path)
get_working_directory	Get current working directory
set_working_directory	Set working directory (params: path)

Edit Tools

Tool	Description
insert_at_line	Insert at specific line (params: file_path, line, content)
replace_lines	Replace line range (params: file_path, start_line, end_line, content)
append_to_file	Append to file (params: file_path, content)
prepend_to_file	Prepend to file (params: file_path, content)

Search

Tool	Description
glob_search	Find files by pattern (params: pattern, cwd)
grep_search	Search file contents with regex (params: pattern, path)
find_definition	Find code definitions (params: name, path)

Command Execution

Tool	Description
execute_command	Run shell commands (params: command, cwd, timeout)
execute_background	Run commands in background (params: command, cwd)
read_output	Read background process output (params: session_id)
kill_session	Kill background process (params: session_id)
list_sessions	List running processes

Git

Tool	Description
git_status	Repository status
git_diff	Show changes
git_log	Commit history
git_add	Stage files
git_commit	Create commits
git_branch	List/create branches
git_checkout	Switch branches
git_push	Push to remote
git_pull	Pull from remote
git_clone	Clone repositories

Web

Tool	Description
web_search	DuckDuckGo search (params: query)
web_image_search	Bing image search with download (params: query)
web_fetch	Fetch webpage content (params: url)
web_fetch_image	Download image from URL (params: url)

Memory & Planning

Tool	Description
memory_store	Store notes with tags
memory_recall	Search notes
memory_list	List all notes
scratchpad_write	Write to scratchpad
scratchpad_read	Read scratchpad
plan_create	Create execution plans (params: goal, steps)
plan_status	Check plan progress
task_add	Add todo items
task_list	List todos

GitHub Blog

Tool	Description
blog_init	Initialize a Jekyll blog for GitHub Pages
blog_post_create	Create a new blog post
blog_page_create	Create a static page
blog_category_create	Create a category page
blog_post_list	List all blog posts
blog_nav_update	Update navigation menu
blog_deploy	Deploy to GitHub Pages
blog_config	Update blog configuration
blog_theme	Apply theme preset or custom colors
blog_theme_list	List available themes
blog_jekyll_theme	Apply a Jekyll remote theme

Skills

Tool	Description
list_skills	List installed skills
load_skill	Load skill instructions
install_skill	Install from GitHub

Utilities

Tool	Description
get_current_time	Current date/time
calculator	Math expressions
think	Record thinking notes
ask_user	Prompt user for input

Skills System

Skills are instruction packages that teach the AI specialized tasks. The model auto-detects which skill to load based on your request.

Installed Skills (16)

Code Quality & Security:

• code-review - Code review methodology
• differential-review - Security-focused diff review
• static-analysis - CodeQL, Semgrep, SARIF
• testing-handbook-skills - Fuzzers, sanitizers, coverage

Web Development:

• react-best-practices - React patterns (Vercel)
• web-design-guidelines - UI/UX fundamentals
• shadcn-ui - Modern component library
• frontend-design - Frontend UI/UX
• web-artifacts-builder - HTML/React prototypes

ComfyUI & Creative:

• comfyui-nodes - Custom node development (V1 + V3 API)
• comfyui-workflow - Workflow creation (SD1.5/SDXL/SD3.5/Flux)

Development Tools:

• chrome-extension - Chrome extension development (MV3)
• mcp-builder - Build MCP servers
• modern-python - Python tooling (uv, ruff, pytest)
• docx - Word document creation
• github-blog - Jekyll blogs for GitHub Pages

Install More Skills

install-skill.bat https://github.com/anthropics/skills/tree/main/skills/pptx

Using Skills

Skills auto-load when they match your request. You can also load them manually:

"What skills do I have?"
"Load the comfyui-nodes skill"
"Use the docx skill to create a report"

Configuration

HTTP Server Port

Edit frontend/server.js:

const PORT = 3847;

Environment Variables

Variable	Default	Description
IMAGE_DOWNLOAD_DIR	~/lmstudio-images	Image download location

LM Studio Setup (HTTP Mode)

1. Load a model (Qwen3 recommended)
2. Enable API server (default: localhost:1234)
3. Open http://localhost:3847/chat.html
4. Enter API token if authentication is enabled

LM Studio Setup (MCP Mode)

1. Add MCP config (see Quick Start above)
2. Restart LM Studio
3. Tools appear automatically in model context

API Endpoints (HTTP Mode)

Endpoint	Method	Description
/	GET	Image viewer UI
/chat.html	GET	Chat interface
/tool	POST	Execute tool {name, args}
/skills	GET	List skills (JSON)
/skill?name=X	GET	Get skill details

Testing

# Test a tool
curl -X POST http://localhost:3847/tool \
  -H "Content-Type: application/json" \
  -d '{"name":"get_current_time","args":{}}'

# List skills
curl http://localhost:3847/skills

Troubleshooting

Server Issues

# Check if running
curl http://localhost:3847/skills

# Restart
restart-chat.bat

# Force stop
stop-chat.bat

MCP Connection Issues

1. Check LM Studio console for errors
2. Verify path in mcp-servers.json is correct
3. Run node src/index.js manually to test

read_file Timeouts / WebSocket Errors

The read_file tool defaults to 500 lines to prevent LM Studio WebSocket timeouts. For large files, use pagination:

read_file with offset=1, limit=100    # Lines 1-100
read_file with offset=101, limit=100  # Lines 101-200

Tool Name Errors

If the model calls wrong tool names (e.g., edit instead of edit_file), the server has built-in aliases that route common mistakes. If you still see errors, update the system prompt in LM Studio — see SYSTEM_PROMPT.md for the correct prompt.

Tool Errors

1. Check browser console (F12) for HTTP mode
2. Verify working directory permissions
3. Check LM Studio model supports function calling

License

MIT