Herding vLLM subprocesses, containers, and remotes from one browser tab

What it does

vLLM Playground is a FastAPI and JavaScript web UI that wraps the vLLM inference engine. It lets you start, stop, and switch between multiple vLLM backends—subprocess, container, or remote—through browser tabs, then chat with them, run vision models, or benchmark throughput. The project also folds in extras like multimodal generation, tool calling, and an observability dashboard.

The interesting bit

The most unusual trick is the Claude Code integration: it embeds a web terminal so you can point Anthropic’s coding agent at open-source models served by your own vLLM instance. Add in MCP (Model Context Protocol) support for agentic tool use with human approval, and the UI starts to feel less like a chat wrapper and more like a local AI ops console.

Key highlights

• Multi-instance management — Run subprocess, container, and remote vLLM servers in parallel, switching between them via tabs and a registry-backed Instances page.
• Multimodal and vision support — Chat with VLMs such as Qwen2.5-VL and LLaVA, and generate audio or images through vLLM-Omni integration.
• Claude Code backend — Use open-source models as the engine for Anthropic’s Claude Code through an embedded terminal interface.
• Enterprise deployment traits — Runs on macOS Apple Silicon (CPU and Metal GPU), and ships with OpenShift/Kubernetes manifests for dynamic pod creation.
• Agentic tooling — MCP server integration with presets for filesystem, Git, fetch, and time tools, plus function calling for Llama, Mistral, and Qwen.

Verdict

Developers already running (or struggling to juggle) multiple vLLM instances will get the most mileage; if you just need a single local chat UI, this is probably overkill.