mcpworld

# Tool List **English** | [English](README_en.md)     > ### 🌐 Use local MCP with ChatGPT Web Browser > MCP is commonly associated with desktop clients like Claude Desktop, but > **MCP World** exposes a local MCP server to a public URL via **SSH reverse tunneling** > and connects it directly to the **ChatGPT Developer Mode Connector**. > This allows for controlling Blender, CAD, Photoshop, HWP, and Office naturally > through the ChatGPT web interface without requiring a separate desktop app.  A GUI for toggling local MCP servers and tunnels connected to the ChatGPT web interface. > ⚠️ **Windows 11 only**. Not compatible with macOS or Linux. > This is due to the dependency on **Windows COM automation** for controlled apps > (HWP, MS Office, Photoshop, CAD) and hardcoded execution paths in `config.json`. ## Initial Setup (One-time) 1. Copy `config.example.json` to `config.json`. 2. Fill in values according to your environment: - `vps_ssh` — VPS SSH target (e.g., `root@YOUR_VPS_IP`) - `public_base` — Public domain (e.g., `https://YOUR_DOMAIN`) - `cwd`, `server`, and `install` paths for each MCP 3. `config.json` is included in `.gitignore` to protect personal settings. > Refer to the template under **"Add a new MCP to VPS"** for VPS nginx routing. > ℹ️ All 5 MCPs are automatically installed from **public repositories** > (github.com/JS190-prog/*) or PyPI (`uvx`). `config.example.json` links to actual > repositories, so **fill in VPS and domain values** and they will be cloned and > installed automatically at startup. To customize, fork each repository and > change the `install` URL. ## Configuration - `mcpworld.pyw` — GUI main body (tkinter + pystray tray) - `config.json` — MCP definitions (ports, paths, execution commands, environment variables) - `.venv\` — Dedicated virtual environment (pystray, pillow) - `logs\` — Server and tunnel logs for each MCP (`<id>-server.log`, `<id>-tunnel.log`) ## Managed MCPs | MCP | Local Port | VPS Port | Public URL | |---|---|---|---| | Blender | 18000 | 8010 | https://YOUR_DOMAIN/bmcp/mcp | | CAD | 18001 | 8011 | https://YOUR_DOMAIN/cmcp/mcp | | Photoshop | 8002 | 8012 | https://YOUR_DOMAIN/pmcp/mcp | | HWP 2024 | 18004 | 8014 | https://YOUR_DOMAIN/hmcp/mcp | | Office | 18005 | 8015 | https://YOUR_DOMAIN/omcp/mcp | Each row: **Server** (local port) + **Tunnel** (SSH reverse tunnel) status indicators → both green indicates readiness for ChatGPT usage. ## Execution - Double-click the **`MCP World`** shortcut on the desktop, or - `\.venv\Scripts\pythonw.exe mcpworld.pyw` ## Usage 1. **▶ Start All** — Start all servers → wait for ports → establish reverse tunnels 2. When all MCP status indicators (server and tunnel) are green, it's ready 3. Run target apps (Blender / ZWCAD / Photoshop / HWP / Office) and open working documents - HWP and Office can be auto-launched via `env` settings in `config.json` 4. Add the public URL to the ChatGPT Developer Mode Connector (no authentication) 5. When finished, **■ Stop All** > **⚙ Settings** button allows editing of all MCP ports, paths, environment variables, > and server commands in the GUI (changes applied after restart). --- ## CAD File Upload Notes Files uploaded to the ChatGPT web interface have a `/mnt/data/...` path, which is an internal path in the OpenAI sandbox. Local Windows PC CAD MCP servers cannot directly access this path. Therefore, the following method is not considered automatic processing: ```json { "file_path": "/mnt/data/plan.dwg" } ``` Previously, the file name was used to find and open a matching file in `C:\cad-mcp\workspace`, `C:\cad-mcp\inbox`, or `Downloads`, but this approach was removed due to potential issues with incorrect file opening. Safe paths allowed by CAD MCP: - Users directly open the original drawing in CAD and use current document tools like `get_active_document_info`, `list_text_entities`, `update_entity`, `save_as_generated`. - File bytes transmitted via `upload_and_open_drawing(file_base64=...)` or `upload_drawing_chunk` are treated as uploaded files. - Only files in explicit local Windows paths or `C:\cad-mcp\inbox` are opened. `file_id` also cannot be automatically retrieved by the local server. It is rejected unless a separate bridge sends the actual file bytes. --- ## Adding a New MCP to VPS (How-To) To use a new tool (MCP) with ChatGPT, follow these 3 steps: **① Local registration → ② VPS exposure (nginx) → ③ ChatGPT connection**. **Connection flow** ``` ChatGPT ──HTTPS──▶ https://YOUR_DOMAIN/<path>/mcp │ (VPS nginx reverse proxy) ▼ VPS 127.0.0.1:<vps_port> │ (SSH reverse tunnel — auto-generated by GUI) ▼ Local PC 127.0.0.1:<local_port> │ (local MCP server) ▼ Target App (Blender / HWP / Office ...) ``` ### Step 1 — Local Registration (`config.json`) Add a new item to the `mcps` array in `config.json` or use the GUI's **⚙ Settings** button: ```json { "id": "myapp", "name": "My App", "local_port": 18006, "vps_port": 8016, "path": "/amcp/", "cwd": "C:/path/to/server", "server": ["C:/path/to/python.exe", "server.py", "--port", "18006"], "env": { "MY_FLAG": "1" } } ``` | Field | Meaning | |---|---| | `id` | Unique identifier (used in log file names) | | `local_port` | Available local port | | `vps_port` | Available VPS port (avoid conflicts in the 8010~ range) | | `path` | Public path (e.g., `/amcp/`) → URL is `.../amcp/mcp` | | `cwd` / `server` | Server execution folder / execution command (array) | | `env` | (Optional) Environment variables to pass to the server (e.g., auto-launch HWP) | > **stdio-based MCPs** are wrapped with `mcp-proxy` for HTTP exposure (used by CAD, HWP, Office). > Example: > ``` > "server": ["C:/cad-mcp/.proxy-venv/Scripts/mcp-proxy.exe", > "--host", "127.0.0.1", "--port", "18006", > "--transport", "streamablehttp", "--pass-environment", "--", > "C:/path/python.exe", "server.py"] > ``` Restart the GUI → start the new row → if the server status indicators are green, local registration is complete. ### Step 2 — VPS Exposure (nginx Routing) Connect to VPS via SSH and add a `location` block to the nginx configuration file. ```bash ssh root@YOUR_VPS_IP nano /etc/nginx/sites-available/automaton-dashboard-http.conf ``` Add the following block — replace **`/amcp/` (path) and `8016` (vps_port)** with your values: ```nginx # === amcp (ChatGPT connector) BEGIN === location ^~ /amcp/ { rewrite ^/amcp/(.*)$ /$1 break; proxy_pass http://127.0.0.1:8016; # ← vps_port proxy_http_version 1.1; proxy_set_header Host $host; proxy_set_header X-Real-IP $remote_addr; proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for; proxy_set_header X-Forwarded-Proto $scheme; proxy_set_header Connection ""; proxy_buffering off; proxy_cache off; proxy_read_timeout 24h; proxy_send_timeout 24h; client_max_body_size 150M; } # === amcp END === ``` Save and **verify before applying** (syntax errors can halt other services): ```bash nginx -t && systemctl reload nginx ``` > **Revert:** Remove the added `# === amcp BEGIN/END ===` block and > `nginx -t && systemctl reload nginx`. Existing blocks are marked with > `# === <name> BEGIN/END ===` comments for easy identification. ### Step 3 — ChatGPT Connection Add the public URL to the ChatGPT Developer Mode Connector (no authentication): ``` https://YOUR_DOMAIN/amcp/mcp ``` - The URL can be copied using the **📋 URL** button in the GUI. - **Tool changes** require deleting and re-adding the connector for the new tool list to appear. --- ## Notes - **Local PC must be on** and server + tunnel must be alive for ChatGPT to work. - Clicking the window X button → minimizes to tray (continues running). Tray icon → exit. - "Exit (server maintained)" closes the GUI but keeps servers and tunnels running in the background → completely shut down requires **Stop All**. - Recommended to enable only when in use due to unauthenticated public exposure.