Sampling and Elicitation are the parts usually missing from desktop clients, and they are what let a server ask the model a question or ask you one mid-call — which is a different class of server than tools alone can support.
An Electron chat application built around MCP, designed to orchestrate several LLM APIs from different vendors through one configuration. You download it and run it; there is no sign-up.
- Coverage of the MCP surface beyond tools: server Tools, Prompts and Resources, plus client-side Sampling and Elicitation
- Real-time MCP server discovery against the MCP registry, and support for MCP Bundles (`.mcpb`), the format previously called Desktop Extensions (`.dxt`)
- Cross-vendor LLM orchestration from one JSON config — either a single chatbot object or an array of them, each with its own name, key, base URL, path, model and model list
- Windows, Linux and macOS builds, with multilingual interface support
- Configuration files you can edit before or after packaging: `llm.json` for endpoints, `mcp.json` for servers, plus startup and popup screens; edits are kept in `localStorage` and clearable from the tray menu
An LLM backend that supports tool or function calling — ChatGPT, Claude, Qwen or something self-hosted — configured with its API key and endpoint. What else you need depends on the servers you run: Node.js for NPX-based servers, Python and `uv` for UVX-based ones, Docker for containerized ones. On macOS and Linux the default MCP configuration usually needs adjusting for CLI paths or permissions, and spawn `ENOENT` errors are the common symptom; running the server command yourself first and using an absolute path is the documented fix. Remote MCP servers with full OAuth support are reached by adding `mcp-remote` as the command in `mcp.json`.
One command plus a key — npx -y mcp-remote https://YOURDOMAIN.com/sse, then supply credentials