janus

janus

janus is a single Go binary from Vibra-Ingenn that loads a GGUF model on the local machine and serves it over HTTP. Inference uses llama.cpp through Vulkan (AMD, Intel, or NVIDIA) or a CPU fallback. The process also ships a bundled web UI (Assistant, Chat, Kernel, Config, Memory, Skills) and a ReAct tool loop that can read and write files, run shell commands, extract text from documents, and render PDF or Word output.There are no command-line flags. Startup reads a `.env` file (walking from the current directory toward the filesystem root, then next to the executable) and environment variables. The default listen address is 127.0.0.1:8990. The OpenAI-style base URL is `http://127.0.0.1:8990/v1`. Useful paths include `/health`, `/v1/models`, `/v1/chat/completions` (streaming supported), `/v1/tools/list`, `/v1/tools/call`, `/kernel/run`, and `/upload` (50 MB multipart cap).On Linux, `./build.sh` fetches llama.cpp Vulkan `.so` files, compiles `dist/janus`, copies the libraries beside the binary, and sets `RPATH` to `$ORIGIN` when patchelf is available. Otherwise run with `LDLIBRARYPATH` pointing at the directory that contains `libllama.so`. The companion modelget binary in the same repository downloads a GGUF file from Hugging Face into `models/`.Hot-swap a model from the web UI Config tab or `POST /models/load` without restarting. Optional backends besides local GGUF are Ollama (`INFERENCE_BACKEND=ollama`) and cloud routers such as OpenRouter when the matching API key is set.

Copy `.env.example` to `.env` in the directory you launch janus from, or export the same names in the environment. Relative `JANUSMODELPATH` values are resolved against the current working directory, not the source tree.INFERENCE_BACKEND

`vulkan` (default), `cpu`, `ollama`, or `openrouter`.

Path to the `.gguf` file. Required for local Vulkan/CPU inference.

Path to `libllama.so` / `llama.dll`. Auto-detected next to the binary when unset.

Layers to offload to the GPU. `-1` is all layers; `0` is CPU only.

Internal VRAM budget hint in MiB (default 9216).

Context window in tokens.

Maximum tokens per reply (default 4096).

Prompt template: `chatml`, `llama2`, or `alpaca`.

Bind address (default 127.0.0.1:8990). A bare port is treated as `127.0.0.1:PORT`.

Set to any non-empty value to skip the automatic browser launch.

When `true`, require login on admin routes (Config, model load, and similar).

Admin password when auth is on. Generated on first run if unset.

When `true`, block shell commands from the `run_command` tool.

Kernel tool loop: `yolo` auto-runs tools; `safe` prompts before destructive tools.

Used when `INFERENCE_BACKEND=ollama` (default URL `http://127.0.0.1:11434`).

This janus is the Vibra-Ingenn GGUF runner. Distro packages named janus are almost always the unrelated Meetecho Janus WebRTC gateway. Several other projects also ship a `janus` binary.The process takes no flags; a missing or wrong `.env` is the usual startup failure. Local Vulkan/CPU mode refuses to start without a readable `JANUSMODELPATH`. Linux needs `libllama.so` beside the binary or on `LDLIBRARYPATH`. First model load commonly takes 10–60 seconds. The built-in browser opener uses Windows `cmd /c start` and does not open a browser on Linux. Built-in tools can run arbitrary shell commands unless JANUS_SAFE_MODE is set. OCR tools need tesseract on `PATH`. Default bind is loopback; do not expose the port without JANUS_AUTH.

Janus is developed by Vibra-Ingenn as a MIT-licensed Go binary that wraps llama.cpp inference, an OpenAI-compatible router, and a local tool-using assistant. Windows is the primary target; Linux and macOS builds are supported from the same tree.

ollama(1), llama-cli(1), llama.cpp(1), llamafile(1), koboldcpp(1), vllm(1), tesseract(1)

Source code · Documentation