lucascs@proton.me Lucas C.S. LocalAI serves local models over an OpenAI-compatible HTTP API, so any client written against that API can talk to a model on this machine instead of a hosted service. It also ships a browser UI, compiled into the binary, for browsing the model gallery and chatting without a separate client. No acceleration backend is compiled in, and that is upstream's design rather than a limitation of this ebuild: LocalAI resolves its inference backends — llama.cpp, whisper.cpp, diffusers, and the rest — as OCI artifacts fetched at RUNTIME, choosing one that matches the hardware it finds. A CPU-only install is therefore the default and needs nothing extra, and nobody has to pull in CUDA or ROCm to run this package. That is why this package has no USE flags and no accelerator dependencies. The build reaches no network. Upstream's Makefile makes three fetches: it downloads a prebuilt protoc, runs `go install ...@version` twice for the two protobuf code generators, and runs `npm install` for the React UI. All three are replaced here. protoc comes from dev-libs/protobuf. protoc-gen-go is built from LocalAI's own vendored sources, so the generator matches the protobuf runtime the binary links against, and protoc-gen-go-grpc — which is a separate Go module and so never reachable through vendoring — is built from a small bundle of its own. The React UI's node_modules and the Go dependencies both arrive as distfiles, so every input to the build is recorded in the Manifest. mudler/LocalAI https://github.com/mudler/LocalAI/issues https://localai.io/docs/ https://github.com/mudler/LocalAI/releases