Your model.
Your machine.
Your code.
Lovanix is a build harness that runs the model on your own GPU and shows you every change before it touches disk. No API key. No per-token meter. No file written behind your back.
Cost per run: $0. The only bill is the electricity you were already paying.
Lovanix wants to write a file
create index.html · new file · 7 lines
@@ -1,1 +1,7 @@-+<!DOCTYPE html>+<html>+<body>+ <h1>Lovanix</h1>+ <p>Built on this machine, for nothing.</p>+</body>+</html>
Nothing happens until you say so
Every tool call stops at a gate and shows its work. Reject one with a reason and the model is told what you said, so the next attempt differs rather than repeats.
Point it at a folder
Lovanix can read and write inside that folder and nowhere else. Path escapes are refused rather than sanitised, and the refusal shows up in the transcript where you can see it.
Ask for the thing you want built
The model runs on your machine through Ollama. Attach files, photos, or any of the skills already installed for Claude Code and Codex. They are read in place, from the same directories, so there is one copy of the truth.
Approve, or reject with a reason
A new file is diffed against empty, so you read the whole thing before it exists. Always allow is scoped to this session and this folder, and a shell allowance covers the exact command, never the tool.
Watch it render
The preview pane serves the project over its own scheme, sandboxed and confined by the same check the file tools use. Save a file, the pane reloads.
The model has to call tools, not describe them
Measured on this machine, an RTX 5090 with 31 GB of VRAM, against the same one-file task. The reasoning distills talk themselves through the work and then narrate the action instead of emitting a call. A 4B model trained for tool use beats a 32B model that was not, and it is not close.
| Model | Called the tool | Time to finish |
|---|---|---|
| qwen2.5:7b-instruct | yes | 0.3 s |
| qwen3:4b | yes | ~4 s |
| deepseek-r1:14b | no, narrated it | ~2 s |
| deepseek-r1:32b | no result at all | killed at 30 min |
| llama3.2:3b | no, narrated it | ~2 s |
Lovanix sizes the default to your card and will not hand you a model from a family measured as narrating, at any amount of VRAM.
Get it running
Three commands and a folder. Everything below runs offline once the weights are pulled.
# 1 — install Ollama, then keep it serving $ ollama serve # 2 — pull a model that can call tools $ ollama pull qwen3:4b # 3 — check the machine end to end $ npm run doctor # 4 — open it $ npm run dev
Set OLLAMA_MODELS first if you want the weights off your system drive. Ollama does not move them afterwards.
The doctor is the real gate. It checks Ollama, the models, your VRAM, a live reply, and one real tool call, in that order, so a failure names its own cause.
Nothing leaves the machine. No telemetry, no account, no cloud fallback. If Ollama is down, Lovanix says so and blocks Send rather than looking fine and failing later.
Your skills come with you. Anything already installed under .claude, .codex, or .agents is listed and runnable, read in place and never copied.
Where this actually stands
Lovanix drives a real local model through a multi-step task, survives a restart with the folder, the model and the conversation intact, fails legibly when Ollama is not running, and stops mid-run when you tell it to. All four were watched end to end on real hardware, not inferred from tests.
What is not done: the installers are not code-signed yet, so Windows and macOS will warn on first launch. Run it from source until that changes. Connectors and plugins are visible but disabled. There is no MCP client behind them, and a live-looking button for something that does nothing is the wrong lie to tell.