Local inference

Your models.
Your way.

Revolver is an open source control plane for running AI on your devices. Download models (GGUF, MLX, safetensors) in app or use your own.

Downloads Dev log
GGUF MLX on macOS Native llama-server Optional Docker OpenAI-compatible gateway

What it is

Chat with your loaded models with a simple, lightweight chat interface. Load models with a setup server wizard. Download models. No telemetry. No metrics. Lightweight.

Chat

Talk to a running server

Multi-turn chat with markdown, KaTeX math, and reasoning traces when the model emits them.

Models

Local files and Hugging Face

Installed GGUF on disk, plus hub search and download. The hub token is stored in settings.

Servers

One process per instance

Create, load, and stop servers. Pin a GPU. New instances default to a native host process. Docker remains available.

Config

Paths, VRAM, and gateway

Default context length, GPU layers, and KV cache. An OpenAI-compatible endpoint for other clients.

Monitor

Host and GPU usage

CPU, memory, disk, and swap. GPU and VRAM when the driver stack reports them.

Runtimes

Engines on first use

The Electron shell is thin. llama.cpp and MLX download from GitHub releases, with a checksum, into the data directory.

The interface

Screenshots from the current desktop app.

Chat tab: conversation against a running Gemma server
Chat — sessions, model picker, and streaming replies.
Models tab: installed GGUF and Hugging Face search
Models — installed weights and the Hugging Face hub.
Server list with one native llama.cpp instance running
Servers — instance list, status, and new server.
Server detail with gateway URL and llama-server logs
Server detail — gateway URL, start and stop, live logs.
Config tab: load defaults, VRAM guardrails, runtime SKUs
Config — load defaults, VRAM limits, and runtime SKUs.
Monitor tab showing CPU, memory, disk, and swap
Monitor — host meters, and GPU when drivers report it.

How it runs

The current product is the Electron native app. Docker Compose is still available when you want a browser UI and container-only inference. Notes are in the dev log.

Electron, native

Desktop app. The main process hosts the control plane. New servers start a host llama-server. Runtimes download on first use. This is the default path.

Docker Compose

nginx serves the UI on port 8080. The backend listens on loopback. Inference runs in containers through docker.sock. Native spawn is disabled.