Work
The runtimes, tools, and experiments behind Caletta. Each project page describes its current availability and scope.
Start with running a model, working with CUDA, or isolating an agent run. For the latest engineering work, read the build notes.
Model runtimes and GPU tooling
mlx-goArrays, autograd, and the MLX runtime in Go.
mlx-go-lmModel loading, generation, serving, embeddings, and fine-tuning.
ModelIRInspect model structure and generate supported implementations.
lib/nvidiaCUDA bindings, profiling, diagnostics, and regression trials.
gputraceInspect and compare Apple Metal GPU captures.
Apple frameworksGo bindings for Apple platform APIs.
Agent environments and tools
coveFork macOS guests for agent runs and CI.
skiffRun agent work in isolated container environments.
axmcpExpose macOS accessibility interactions to agents.
cdpBrowser control through the Chrome DevTools Protocol.
nlmNotebookLM tooling.
AI StudioTooling for Google AI Studio workflows.
Routing, training, and distributed experiments
llm-routerRoute model requests and retain consented examples.
mlx-meshDistributed training research across Macs and CUDA devices.
mlx-go-cclCollective communication for MLX workloads.
mlx-go-irohPeer connectivity for distributed model work.
mlx-go-fedistillFederated distillation experiments.
mlx-go-tallytrainTraining contribution accounting.
mlx-go-siaSampled inference verification experiments.
mlx-go-aneApple Neural Engine experiments.
mlx-go-vibethinkerTraining and adaptation experiments.
OpenTinkerAn open training-service interface.