A traffic controller for your menagerie of local LLM engines

What it does vLLM Studio is a local-first workstation that acts as a control plane and orchestration layer for self-hosted LLM inference. It wraps vLLM, SGLang, llama.cpp, MLX, and exllamav3 behind a single Bun/Hono controller that handles model lifecycle, GPU metrics, logs, and an OpenAI-compatible proxy. A Next.js frontend—usable as a web app or an Electron desktop shell—and a small CLI let you launch models, chat with endpoints, and run agent sessions against a local or remote controller.

The interesting bit The controller owns runtime target discovery and process coordination, surfacing everything from audio endpoints to server-sent events through one API. Rather than spawning a separate agent process, the frontend loads @earendil-works/pi-coding-agent directly in its own runtime, which keeps the architecture flat and lets you point a laptop frontend at a remote GPU controller instead of living inside an SSH session.

Key highlights

• Normalizes five distinct backend families behind a single OpenAI-compatible proxy.
• Ships as both a Next.js web UI and an Electron desktop app, plus a Bun CLI for terminal operators.
• Built-in agent surface using @earendil-works/pi-coding-agent loaded directly in the frontend runtime.
• Controller can run locally or remotely, with frontend and CLI able to target either.
• Surfaces GPU metrics, logs, usage data, and runtime events via server-sent events.

Caveats

• It does not bundle the inference engines; you still need to bring your own CUDA stack, Python environments, or llama-server binaries.
• macOS or Linux only for local development, and the prerequisite list spans Node.js, npm, Bun, and optional GPU drivers.
• Remote deployment relies on SSH and manual .env.local configuration, so it is not a turnkey managed service.

Verdict Worth a look if you self-host multiple LLM backends and want one dashboard to rule them all. Skip it if you expect a batteries-included installer or a managed cloud API.