Janus: run local AI models on your PC, managed from one console

Your models are scattered: some in LM Studio, some in Ollama, a few in Downloads. Every app wants its own setup. Janus finds them all, runs the one that fits your graphics card, and gives every app one address to use.

One program, no Python, no Docker. Double-click it and the console opens in your browser.

Free and open source (MIT).

---

What's new

• The console. A control panel, not a chatbot. Five screens:
  • My Models: every model file on your PC, its size, and whether it fits your graphics card. Load and unload with one click.
  • Discover: search Hugging Face, see which file fits your PC, download it with a progress bar.
  • Server: the address to copy, ready-made examples for your apps, and a live log of every request.
  • Test: send one prompt and see the reply, the speed and the memory used.
  • Settings: every setting Janus has, saved from the page. No .env editing.
• Finds the models you already have: Janus's own models\ folder, LM Studio, Ollama, the Hugging Face cache, and the rest of your drives (it scans once on first start). Add any other folder from the page.
• Picks for you. On first start it loads the biggest model that fits your graphics card (or your RAM, without one).
• Fits big models on your card. If a model and its working memory don't fit together, Janus uses a smaller working memory instead of failing, and falls back to the processor only as a last resort.
• Thinking models just work. The thinking of Qwen 3.x, DeepSeek-R1, Gemma 4 and gpt-oss goes in a separate reasoning_content field, so your app gets a clean answer.
• Local and cloud behind one address. Model names like openrouter/…, gemini/… or claude-… go to those services with your own key; a fallback list takes over when one fails. "Use only this PC" switches all cloud models off.
• Double-click start. The browser opens by itself. Double-click again while it's running and it just opens the page.
• Works alongside other tools. Janus publishes a small "tool card" at /tool.json, so other tools on your PC can find it and use it.

---

Quick start (Windows)

1. Download the latest Janus-windows.zip from Releases.
2. Unzip it anywhere.
3. Double-click janus.exe. Your browser opens the console at http://127.0.0.1:8990.
4. Pick a model in My Models, or get one in Discover.

To stop Janus, close its window. Closing the browser tab keeps it running for your apps.

---

Connect your app

Setting	Value
Base URL	http://127.0.0.1:8990/v1
API key	leave blank (or any text, if the app insists)
Model	local

Works with anything that speaks the OpenAI API: Cursor, Cline, VS Code extensions, Open WebUI, scripts.

curl

curl http://127.0.0.1:8990/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"local","messages":[{"role":"user","content":"Hello!"}]}'

Python (openai package)

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8990/v1", api_key="none")
reply = client.chat.completions.create(
    model="local",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(reply.choices[0].message.content)

Streaming works too: add "stream": true (or stream=True in Python).

---

API

Method	Path	What it does
POST	/v1/chat/completions	Chat (streaming supported)
GET	/v1/models	Model list
GET	/models/list	Every model file Janus found
POST	/models/load	Switch model without restarting
POST	/models/unload	Free the graphics memory
GET	/engine/status	Loaded model, backend, memory used
GET	/health	{ok, id, version}
GET	/version	Version, build time and commit
GET	/tool.json	Janus's tool card

---

Settings without the console

Everything in Settings can also be set as an environment variable or in a .env file next to janus.exe. The most used:

Variable	Default	Meaning
JANUS_MODEL_PATH	picked for you	A specific .gguf file to load
JANUS_MODEL_SEARCH_PATHS		Extra folders to search, separated by ;
INFERENCE_BACKEND	vulkan	vulkan or cpu
JANUS_GPU_LAYERS	-1	-1 = all layers on the graphics card, 0 = CPU only
JANUS_CTX_SIZE	16384	Working memory (context) size; Janus steps down to 8192, then 4096, if it doesn't fit
JANUS_MAX_TOKENS	until the context is full	Longest reply
JANUS_LISTEN_ADDR	127.0.0.1:8990	Address and port
JANUS_NO_BROWSER		Set to 1 to stop the browser opening at start

---

Build from source

Windows (Go 1.25+):

git clone https://github.com/Vibra-Ingenn/Janus.git
cd Janus
.\build.ps1
.\dist\janus.exe

build.ps1 downloads the llama.cpp Vulkan files and builds dist\janus.exe. After that, Rebuild Janus.bat closes the running copy, rebuilds and starts the new one.

Linux (Go 1.25+, Vulkan or CPU):

git clone https://github.com/Vibra-Ingenn/Janus.git
cd Janus
./build.sh
./dist/janus

• build.sh pins llama.cpp b11146 (the struct layouts in the bridge match that tag); ./build.sh --version bNNNN tries another, --skip-download reuses lib/linux/.
• The .so files sit next to the binary and find each other through RUNPATH=$ORIGIN; no LD_LIBRARY_PATH needed.
• Supported: linux/amd64. macOS and arm64 are untested.

---

Things that trip people up

Problem	What's going on	Fix
A rebuild "didn't take"	Windows can't replace a running .exe.	Use Rebuild Janus.bat, or close Janus's window first.
A Janus you can't close	It was started from an admin window; normal windows can't close admin programs.	Run Janus from a normal window. To clear one, close it from an admin window.
"Port is in use"	Another program holds port 8990. Janus names it in its window.	Close that program, or change the port in Settings.
First reply is slow	The model is loading into memory.	Normal: 10-60 seconds. Smaller files (Q4) load faster.
"Engine failed to start"	Graphics driver problem.	Update the driver, or set the backend to CPU in Settings.

---

Works better together

Janus runs fine on its own. It's also one of a set of free tools that find each other through their tool cards:

• Terminus: turns messy PDFs and receipts into clean data.
• Vox Vena: rewrites AI-sounding text so it reads like a person wrote it.
• Vibe Engine: the paid tool that runs them all: an AI writes a job down once, and it runs the same way every time on your PC.

---

Thanks

To thalisantunes for the Linux support (#2). Contributions are welcome; see CONTRIBUTING.md.

Like it?

Janus is free and stays free. If it saves you time, a star on GitHub helps other people find it. Using an AI assistant? Ask your AI to check out adeptuscamini.com.

License

MIT, see LICENSE.

Recusamus pergere eadem fracta facere; nos alia facimus.
_{We refuse to keep doing the same broken things; we do other things.}