MCP Tool Builder
hanwegLABSCO SUMMARY
An MCP host that is useful before you have installed a single server.
Fetch, file access and shell execution are built in, so the first conversation already has tools; the installer agent is what gets you from there to the rest of the ecosystem without editing JSON by hand.
What it is
An open-source desktop application that hosts MCP servers and connects them to the model of your choice. It ships in two builds — a Tauri version with a smaller installer and a traditional Electron version.
What you get
- MCP servers over both stdio and SSE, with per-tool enable and disable rather than all-or-nothing per server
- Three built-in local tools available immediately: Fetch for web requests, File Manager for reading and writing files, and Bash for command execution
- An MCP Server Installer Agent that installs and configures new servers for you
- Any LLM with function calling — ChatGPT, Anthropic, Ollama and OpenAI-compatible endpoints — with multiple API keys and model switching through
model_settings.json - One-click access to managed MCP servers through OAP Cloud, which removes the local Python and Docker setup
- Custom system prompts, keyboard shortcuts for rename, settings, reload and new chat, chat draft saving, chat history search, and
@file-path search in the input box - Handling for several MCP elicitation requests at once, plus MCP server authentication
- 24+ interface languages and an auto-update mechanism
Requirements
Download a build from the releases page. On Windows the Python and Node.js environments are downloaded automatically after launch; on macOS and Linux you install Python and Node.js (with npx and uvx) yourself. Linux users should pick the Electron AppImage, chmod +x it, and on Ubuntu or Debian may need --no-sandbox or a system settings change; Arch users can install dive-ai from the AUR. macOS currently has the Electron .dmg only. MCP server authentication is noted as unstable and may need re-authorizing often.
Setup effort
Follow the published steps
USE CASES
01
Getting started with MCP on a desktop
The bundled Fetch, File Manager and Bash tools work on first launch, so there is something to try before any server configuration exists.
02
Adding a server without editing config files
Ask the built-in installer agent for the server you want and let it handle the install and the configuration.
03
Running everything against a local model
Point it at Ollama or another OpenAI-compatible endpoint; function calling is the only requirement on the model side.
04
Cutting a noisy server down to size
Enable and disable individual tools rather than the whole server, so one over-broad server does not crowd out everything else.
05
Skipping local dependency setup entirely
Sign up at OAPHub.ai and connect managed servers by deep link or config file — no Python, Docker or local runtime to install.
BEFORE YOU INSTALL
What to know first
Requires
Python and Node.js (with npx and uvx) on macOS and Linux — the Windows build downloads them after its first launch
INSTALLATION
Connect your client
Server setup
Download the latest release for your system from https://github.com/OpenAgentPlatform/Dive/releases/latest — Windows: the Tauri installer (recommended) or the Electron one; macOS: the Electron .dmg; Linux: the Electron .AppImage (recommended) or the Tauri build
On macOS and Linux, install Python and Node.js (with npx and uvx) yourself; on Windows they are downloaded automatically after the first launch
Linux AppImage: make it executable with chmod +x, and on Ubuntu or Debian you may need to add --no-sandbox. On Arch Linux you can install it from the AUR instead, for example paru -S dive-ai
