The workflow it is built for — paste a GitHub URL, say install this in my config — removes the most tedious part of trying new servers. It also means an agent is editing the file that decides what other agents can do. The README's own TODO list names the gap: there is no recovery mechanism for a failed installation yet, and support beyond Claude Desktop is still pending. Keep a copy of the config.
An MCP server whose job is installing MCP servers. It resolves a package from PyPI or npm, or a local directory, works out whether it is a Node or Python server, installs what is needed, and registers the entry in the Claude Desktop configuration file. It runs on Windows, macOS and Linux.
- Installation from a package repository — PyPI or npm — or from a local directory
- Automatic detection of whether the target is a Node.js or Python server, so you do not specify it
- Environment variables and launch arguments configured for the server being installed
- The Claude Desktop config file updated in place, with Node packages wired through `npx` and Python ones as modules
- Cross-platform paths handled for you, including the Windows, macOS and Linux config locations
Python 3.10 or higher and pip; Node.js and npm as well if you want to install JavaScript packages. Installation is from source — clone, then `pip install -e .` — after which the entry point is `mcpbuilder`, or `python -m mcp_builder.server`. The project is `mcp-builder` (0.1.0 in pyproject). It edits the Claude Desktop configuration file directly, so back that file up before letting it run. Recovery from a failed install is listed as not yet implemented.
Build from source — clone the repository and build it, then point your client at the binary