What it is
Lemonade is the open-source local AI server that gives you cloud-style capabilities — chat, coding, speech, image generation — without sending the request out. It serves optimized LLMs from your own GPUs and NPUs.
If Interviewer is the app I point people at, this is the floor it stands on.
Why it matters
Local inference only works if the server is boringly reliable: model loading, streaming, tools, and hardware backends that do not collapse when you leave the happy path. That is the job. During the AMD co-op I treated Lemonade as production infrastructure for local-first agents, not a demo.
How it works
The public repo spans C++, Python, and a TypeScript/React surface. My attributed work there is concentrated on the MLX path: enabling tool_calls capabilities, stream hygiene for tool calls, and the lemon-mlx backend wiring — the kind of glue that lets an agent actually call tools on a local model instead of only chatting.
Status
Actively maintained upstream at lemonade-sdk/lemonade. I am a contributor, not the project lead. Related tooling I touched: lemonade-eval (benchmark dashboard work) and lemon-mlx-engine.
Related
- Lemonade Interviewer — the mock-interview app on this server
- Lemon MLX Engine — C++ MLX backend I also contributed to
- lemonade-eval — benchmarking client; I added dashboard and automation work
- aifaceoff — local head-to-head benchmarks for the pi coding agent