The reason to pick this over writing a server is that the three things a server actually offers — content, prompts and callable functions — are all declarative here, and resource watching means the content stays live. If you already have Python functions worth exposing, pointing at their import paths is a shorter route than wrapping each one by hand.
A server that reads its whole environment from a YAML file. You declare which content the model can see, which prompts it can use, and which Python functions it can call, and the server exposes them over MCP. Building a custom server becomes editing a config file rather than writing an implementation.
- Resources of several kinds — files (`PathResource`), inline text (`TextResource`), the output of a CLI command (`CLIResource`), Python source (`SourceResource`), the result of a Python callable (`CallableResource`), and images (`ImageResource`)
- Resource watching and hot-reload, so a glob like `./src/**/*.py` stays current as the files change, with exclusion patterns
- Tools from your own code — register a Python function by its import path, or pull a whole toolset from an OpenAPI spec; parameters are validated before the call
- Prompts as templates with validated arguments, defined statically, generated from a Python function, or loaded from a file, with completion suggestions for their arguments
- Global settings in the same file: timeout, retry count, log level, extra paths and requirements
- Transports for both cases — stdio by default, SSE or Streamable HTTP for web clients, configurable with host, port and CORS origins
No account and no key for the server itself, though tools you wire in may need their own. Run it with `uvx mcp-server-llmling start path/to/your/config.yml` — that command is the whole client config for Claude Desktop or Zed. Everything else is the YAML: `resources`, `tools`, `toolsets`, `prompts` and `global_settings`. For OpenAPI toolsets, bundling the spec with Redocly CLI first resolves references cleanly, and it is used automatically if installed. There is also a Python API if you would rather start the server yourself. Package `mcp-server-llmling` at 1.0.0.
One command — uvx mcp-server-llmling@latest