Open source · MIT · runs locally

Gemma, GLM and Qwen,
on your machine.

The Tara Vision inference engine: a local, private and OpenAI-compatible server built on llama.cpp, with one pinned build per release, a curated catalog, tuned presets and measured memory. Made for taracode, open to any client. Nothing leaves your machine.

Install taraengine $ taraengine doctor
taraengine - zsh
MIT license100% localzero telemetry8 catalog modelspinned runtime

$ taraengine service install

one command installs the binary and the service that runs the engine. pick your platform:

$ brew install --cask tara-vision/taracode/taraengine
$ curl -fsSL https://engine.tara.vision/install.sh | bash
$ sudo apt install ./taraengine_0.5.0_linux_amd64.deb
$ sudo dnf install ./taraengine_0.5.0_linux_amd64.rpm

apt on debian and ubuntu, dnf on fedora and rhel: each also installs the openmp library that llama.cpp needs.

$ BASE=https://github.com/tara-vision/taraengine/releases/latest/download
$ BIN=taraengine-$(uname -s | tr '[:upper:]' '[:lower:]')-$(uname -m | sed 's/x86_64/amd64/; s/aarch64/arm64/')
$ curl -fsSL --fail-early -O $BASE/$BIN -O $BASE/checksums.txt -O $BASE/checksums.txt.sigstore.json && \
cosign verify-blob --bundle checksums.txt.sigstore.json \
--certificate-identity-regexp 'https://github.com/tara-vision/taraengine/.github/workflows/release.yml@refs/tags/v.*' \
--certificate-oidc-issuer https://token.actions.githubusercontent.com checksums.txt && \
grep " $BIN\$" checksums.txt > $BIN.sha256 && sha256sum -c $BIN.sha256 && \
sudo mkdir -p /usr/local/bin && sudo install -m 0755 $BIN /usr/local/bin/taraengine

on linux, first sudo apt-get install libgomp1 (or sudo dnf install libgomp), the OpenMP library that llama.cpp needs. one chain: a download, a signature or a checksum that fails stops it before anything is installed. then the service: taraengine service install on macos, sudo /usr/local/bin/taraengine service install on linux.

then: taraengine pull gemma4:12b. the service fetches the pinned llama.cpp build for this machine on its own; on a mac the first command starts it. the documentation has every way to install.

$ taraengine doctor

the engine owns everything above the kernels; the math is ggml through llama.cpp

taraengine runtime install

Pinned runtime

One llama.cpp build per release, hashed against the lock before it is unpacked and checked again every time the engine starts.

✓ b11514 verified

taraengine pull

Curated catalog

Models pinned by size and sha256, with the projector and the draft model they come with, and a KV cache chosen for each on an RTX 5090 board.

✓ sha256 verified · 7.2 GB

/v1/chat/completions

OpenAI and Anthropic APIs

llama-server's surface, proxied as it streams: chat, completions, embeddings and messages, with a Bearer key when you set one.

✓ 200 · 248 tok/s

tools, thinking, vision

Tools, thinking and vision

Tool calls, a thinking switch with effort levels, and images where the model reads them.

✓ reasoning_content

taraengine ps

Measured

Every chat answer carries its timings, and ps shows what a loaded model holds in memory, measured on NVIDIA and estimated on a Mac.

✓ 10.0 GiB measured

listen: 127.0.0.1:8272

Privacy-first

Localhost by default and a Bearer key when you set one; no telemetry, no update checks of its own, and the network only for the downloads that a pull, an install or an update needs.

✓ nothing left your machine

taracode

Made for taracode

taracode finds the engine at its address with no configuration, since taracode 3.3.0; any OpenAI-compatible client can use it too.

✓ http://127.0.0.1:8272

$ taraengine doctor --privacy

what leaves your machine: nothing.

No account

No sign-in, no license keys, no waitlist. Install and run.

Zero telemetry

Nothing is collected, ever. Read the data transparency page.

Fully local

Once a model is pulled, the engine works without internet.

$ install taraengine

Install taraengine