Lemonade SDK
by lemonade-sdk
Lemonade serves local LLMs, Whisper, and image models across CPU, GPU, and NPU hardware, and exposes them to any MCP client through a built-in MCP server.
Connect an MCP client
curl -X POST http://localhost:13305/mcp
About
Lemonade is a local inference server: point it at GGUF, FLM, or ONNX model files, including custom ones pulled from Hugging Face, and it runs them on whatever hardware you have — CPU, GPU via Vulkan/ROCm/Metal, or NPU on supported AMD Ryzen AI hardware, with Linux NPU support arriving via the FastFlowLM runtime. Its MCP server turns that local inference into a set of callable tools, so a client like Claude Desktop, GitHub Copilot, Cursor, or the MCP Inspector can drive your local models the same way it would call any cloud API.
That coverage extends beyond text: lemonade_chat handles LLM chat completions with streaming and tool-calling, lemonade_transcribe_audio runs Whisper-based transcription, and lemonade_generate_image produces images with sandboxed disk writes. lemonade_omni goes a step further, orchestrating planning, image generation, editing, and text-to-speech together in one multimodal call, while lemonade_list_models reports what's currently loaded, available, or recommended for your hardware.
Key features
- GGUF, FLM, and ONNX model support across CPU, GPU, and NPU
- MCP server exposes local inference as callable tools for any MCP client
- Whisper-based audio transcription and image generation tools
- Multimodal orchestration tool combining planning, image gen/edit, and TTS
- Linux NPU inference support for AMD Ryzen AI via FastFlowLM
- Model discovery tool reports loaded, available, and hardware-recommended models
Use cases
- Giving an MCP client access to a fully local model instead of a cloud LLM API
- Running Whisper transcription and LLM chat from the same local server behind one MCP endpoint
- Building a multimodal agent flow — plan, generate an image, edit it, narrate it — via a single lemonade_omni call
- Taking advantage of NPU acceleration on AMD Ryzen AI hardware without a cloud dependency
Available tools
lemonade_list_models
Reports which models are currently loaded, available locally, or recommended for your hardware.
lemonade_chat
Runs LLM chat completions, with streaming and tool-call support.
lemonade_transcribe_audio
Transcribes audio using a local Whisper model.
lemonade_generate_image
Generates an image from a prompt, writing output to a sandboxed disk location.
lemonade_omni
Orchestrates a multimodal flow — planning, image generation/editing, and text-to-speech — in one call.
Frequently asked questions
Does Lemonade require a GPU or NPU to run?
No — it runs on CPU by default. GPU acceleration via Vulkan/ROCm/Metal and NPU support via AMD Ryzen AI/FastFlowLM on Linux are available where the hardware supports them.
Which model formats does the MCP server work with?
GGUF is the most portable, since it runs everywhere Lemonade does. FLM and ONNX variants are also supported and can offer better performance on specific hardware.