LocaLM

LocaLM runs AI models on your own hardware for chat and optional tasks such as coding, document questions and media generation. It's for people who want a single local tool and can choose which extra capabilities to add.

• Chat and model choice: Use a browser interface or terminal chat with GGUF models through llama.cpp or HuggingFace Transformers models, including AWQ; an OpenAI-compatible API lets other apps connect.
• Coding agent: An optional plugin edits files, uses shell and search tools, runs tests and provides session diffs for review; you can redirect a task while it runs.
• Document answers and memory: Optional document retrieval grounds answers in your files, while durable memory keeps context across conversations.
• Media and voice: Plugins add image, music and video generation, plus voice interaction and scheduled tasks.
• Model matching and MCP: Requests can use an installed model with the needed image, tool-calling or context capabilities while respecting an explicitly selected model; MCP support connects models to clients such as Claude Desktop and the agent to external tool servers.

LocaLM is open source under AGPL-3.0 and runs on Windows and Linux. It supports CPU inference and NVIDIA, AMD and Intel GPUs through CUDA, ROCm/HIP or Vulkan, depending on the hardware. Local models can work fully offline. Cloud providers are optional, and using them sends requests outside your machine. Chat is built in; the coding agent, document retrieval, memory, media, voice, scheduling and MCP capabilities are optional plugins.