A component library tuned for local-AI apps.
Granola-feel listening. Cursor-feel chat. Linear-feel keyboard.
Every component below is a single hand-crafted React + TypeScript file with a paired CSS file.
Copy them into your repo via locara add <name>, or import directly from
@locara/components. They read from the same design tokens, so they look like one
product from the moment you wire them in.
Composites
2Editors
7// Selection in the editor:
return useMemo(() => data.filter(d => d.active), [data])
The Locara runtime is responsible for honoring the manifest's capability declarations at every system call.
Chat & messaging
4AI affordances
16Local-first AI products typically use shared-memory architectures rather than discrete-GPU systems, because the latency cost of PCIe transfers dominates inference at single-batch.
M3 Ultra features the most powerful CPU and GPU in any Mac, double the Neural Engine cores, and support for up to 512GB of unified memory.
We propose FlashAttention, an IO-aware exact attention algorithm that uses tiling to reduce the number of memory reads/writes between HBM and SRAM.
What would you like to do?
Pick one of the starter prompts, or just start typing below.
Voice & listening
3System & power-user
3The unified memory architecture in Apple Silicon means the CPU, GPU, and Neural Engine all share one pool of DRAM.
Apple Silicon's unified memory means CPU, GPU, and Neural Engine share the same DRAM pool — eliminating the PCIe bottleneck that dominates inference on discrete-GPU systems.
The M3 Ultra has up to 512 GB of unified memory — the highest of any consumer Mac shipped to date, enabling local inference of models in the 70B–405B parameter range at Q4 quantization.