Offering a full web dashboard (local/remote network access) UI for vllm-mlx (AI-coded, open source)

[clickbrain]

EDIT 2: Copilot ran a security and code audit on the system, but it sill has not been thoroughly tested in use. Doing that now.

EDIT: Apologies, Copilot got a little ahead of itself, I have not tested this thoroughly and it posted the discussion before I was ready. Will update once I have it actually tested for remote access and local access. This has also not been tested for any security issues on remote access.

The following was written by Copilot.

Hi everyone,

I wanted to share and offer back a web-based dashboard UI I had built for vllm-mlx using AI tools (GitHub Copilot / Claude). I am not a developer — I use vllm-mlx for personal AI inference on my Apple Silicon Mac and wanted a friendly graphical interface.

What it is:
A Streamlit browser dashboard exposing every function of vllm-mlx through an easy, no-code interface. You can:

• Start / stop / restart the server with connection info shown immediately (base URL, model ID, API key)
• Browse downloaded models and search the entire mlx-community on HuggingFace
• Download new models with one click, filter by bits/size, sort by downloads/likes
• Switch models with one click — optimal settings are loaded from the model card automatically
• Run benchmarks and view historical comparison charts
• Chat with full named conversation history and per-chat model selection
• Control everything remotely from another machine via a management API (port 8502)
• Auto model-switch proxy: your OpenAI client can change the loaded model just by changing the model field
• iFrame embeddable

Repository: https://github.com/clickbrain/vllm-mlx-ui

Disclosure: This was entirely designed and implemented by AI. I directed the work and tested it on my own machine. I am offering it transparently as an AI-authored contribution.

Install for non-developers:

bash <(curl -fsSL https://raw.githubusercontent.com/clickbrain/vllm-mlx-ui/main/install.sh)

I would be happy to open a PR if the maintainers think this would be a good addition to the main repository. Equally happy to keep it as a separate project if you prefer. No pressure either way — just wanted to give back to a project I find incredibly useful.

Thanks for building vllm-mlx!

[clickbrain]

Update — April 2026 (v0.3.57)

A lot has changed since the original post. The biggest change: the entire UI was rewritten from Streamlit to a custom Vue 3 + FastAPI application. This fixed the flickering, lag, and layout limitations of the Streamlit version and enabled a proper reactive UI. The Homebrew tap and install experience also got a full overhaul.

Here's a summary of what the dashboard looks like today:

---

Installation

brew tap clickbrain/vllm-mlx-ui
brew install vllm-mlx-ui
vllm-mlx-ui

Updates are handled through the UI — a notification banner appears when a new version is available and lets you update in one click. No manual brew upgrade needed.

---

What's in the dashboard today

Chat

• Streaming chat with full markdown rendering, code highlighting, image/video attachment (for VLMs)
• Five mode presets: Chat, Code, Creative, Analysis, Precise — each tunes temperature and parameters for that task type
• Optimal button fetches model-card-recommended parameters from HuggingFace for the active mode
• Saved conversations with names, load/delete
• Conversations continue running in the background even if you navigate to another tab

Models

• Library tab: downloaded models, one-click load, delete, Favorites
• Find tab: HuggingFace search with company quick-filter chips (Meta, Qwen, Google, Microsoft, Mistral, Apple, DeepSeek, MLX Community), filter by fit/size/downloads/likes, trending score column, hide-downloaded toggle
• Favorites tab: starred models for quick access

Serve

• Live metrics at the top: tokens/sec, active requests, memory used, memory %, uptime
• Start/stop server, Release Memory (frees GPU/unified-memory allocation without stopping the mgmt server)
• Connection Info panel with local URL, network URL, API key
• Remote access toggle (binds to 0.0.0.0)

Benchmarks

• Live tab: real-time requests-over-time and GPU memory charts with adjustable time range; KV-cache hit rate stats
• Run Tests tab: select models + test suites (Speed, Quality — GSM8K / MMLU / HumanEval, or both), name your run, live streaming log during execution, Stop Run button, combined Speed + Quality results card
• History tab: searchable/filterable list of all past runs; select two runs → Compare; per-run model charts
• Advisor tab: in progress — will recommend the best model/config for your hardware and use-case

Settings

• All server options surfaced in the UI with improved descriptions and documentation links
• GPU Memory Utilization displayed as a %, browser folder picker for SSD KV Cache directory
• All performance-impacting settings (Paged KV Cache, Continuous Batching, quantization, etc.) can be toggled individually before a benchmark run to compare with/without

Docs

• Full documentation bundled inside the installed package (no longer requires a separate repo checkout)
• Left-nav with search, right-side in-page TOC with scroll-spy, heading anchor links for deep-linking to specific sections
• Dashboard User Guide updated to cover all current features

---

Repository: https://github.com/clickbrain/vllm-mlx-ui
Current version: v0.3.57

Still AI-assisted development (GitHub Copilot / Claude), still running on my M4 Max. Happy to answer questions or take feedback. PR offer stands if the maintainers are ever interested.