go-vision-mcp

Vision MCP

A Go-based MCP (Model Context Protocol) server that enables vision analysis for LLMs without native vision capabilities. Supports two backends:

• Local — runs a vision model via llama.cpp entirely on your machine. Private, no data leaves your PC. Auto-downloads models from HuggingFace.
• Gemini API — uses Google Gemini's free tier. No model download, no GPU needed. Just paste an API key from Google AI Studio.

Features

• Nine MCP tools — image analysis from URLs, local files, clipboard, and clipboard history (Windows / Linux with X11 or Wayland)
• Dual backend — choose between fully private local inference or cloud Gemini API (free tier)
• Clipboard history monitor — optionally polls clipboard for images, keeps a configurable history for multi-image comparison (e.g. "the first image is the before, the third is the after")
• Lazy loading — model downloads + llama-server starts only on first tool call, saves bandwidth/VRAM when idle
• Auto-download — downloads models from HuggingFace and llama-server binaries on demand
• Hardware-aware — auto-detects GPU (CUDA/Metal/Vulkan) and recommends optimal quantization
• Automatic resume — interrupted downloads resume from where they left off
• Idle timeout — automatically unloads model from GPU memory after configurable inactivity (default: 5 min), reloads on next tool call
• Multiple clients — concurrent MCP clients (Kilo Code, OpenCode, etc.) share the same llama-server; automatic recovery if another client stops it
• Configurable — supports manual model paths, LM Studio models, and custom llama-server binaries
• TUI wizard — Bubble Tea interactive setup wizard for guided configuration
• Graceful shutdown — kills llama-server on client disconnect (no orphaned processes on Windows)

System Requirements

Requirements depend on the selected model and quantization. The default model is Qwen3-VL-4B-Instruct-GGUF (~4B parameters). The TUI wizard lets you choose between multiple models. The tool automatically detects your hardware and selects the optimal quantization — no manual tuning needed.

Guidance by VRAM (default model: Qwen3-VL-4B-Instruct)

VRAM	Recommended Quantization	Approx. Size	Quality
≥ 12 GB	Q8_0 / Q6_K	3.5 – 4.5 GB	Maximum
≥ 8 GB	Q5_K_M	~3.1 GB	High
≥ 6 GB	Q4_K_M	~2.7 GB	Balanced
≥ 4 GB	Q3_K_M / IQ4_XS	2.1 – 2.5 GB	Economy
≥ 2 GB	UD-IQ3_XXS	~1.8 GB	Ultra low RAM

• RAM: 8 GB+ recommended (4 GB minimum)
• Disk: 5 GB free for model storage
• Network: Required for initial model download
• GPU: NVIDIA (CUDA), AMD/Intel (Vulkan), or Apple Silicon (Metal)

Quick Start

Option A: Local (private, auto-download)

The tool automatically detects your GPU, RAM, and VRAM, then selects the best model and quantization for your system.

1. Download the binary from the latest release — or build it yourself:

go build -o vision-mcp.exe .

2. Run it:

# Quick setup — auto-detects hardware, downloads model, starts server
vision-mcp.exe

# Or use the interactive setup wizard
vision-mcp.exe --configure

Option B: Gemini API (free tier, no download)

Use Google Gemini's free tier — no model download, no GPU needed, no data stored locally.

# Configure with a Gemini API key
vision-mcp.exe --gemini

You will be prompted to paste an API key. Get one for free at:

https://aistudio.google.com/app/api-keys

Click "Create API Key" and copy the generated key. Once saved, run vision-mcp.exe normally to start the MCP server with the Gemini backend.

To switch back to local mode, either run --configure again or edit config.json and set "backend": "local". The API key is preserved in the config so you can switch freely.

CLI Reference

Flag	Description
(none)	Start MCP server (STDIO mode)
--configure	Open interactive TUI setup wizard
--install	Quick non-interactive install with auto-detected defaults
--manual	Manual config wizard (LM Studio, custom paths)
--gemini	Configure Gemini API key (free tier)
--free	Free GPU memory — stop llama-server on port 8001
--status	Show hardware, model, and config status
--download	Download/verify models without starting server
--uninstall	Remove installation directory
--generate-agent-config	Generate markdown with setup instructions for your agent
--version	Show version

Configuration

Configuration is stored at ~/.go-mcp/vision/config.json (Windows: %USERPROFILE%\.go-mcp\vision\config.json).

Fields

Field	Default	Description
repo_id	unsloth/Qwen3-VL-4B-Instruct-GGUF	HuggingFace repo for GGUF model
quantization	UD-IQ3_XXS	Quantization variant to download
mmproj	mmproj-F16.gguf	Multimodal projector filename in the repo
llama_backend	cuda	Compute backend: cuda, cpu, vulkan, metal
llama_bin	llama-server.exe	llama-server binary name
models_dir	~/.go-mcp/models/llm	Directory for downloaded models
port	8001	Port for llama-server
n_ctx	8192	Context size (tokens)
ngl	999	GPU layers (-ngl). 0 = CPU only
flash_attn	true	Enable flash attention (-fa on)
auto_download	true	Download model/mmproj automatically
download_mirror	https://github.com/ggml-org/llama.cpp/releases	Mirror for llama-server binary
custom_prompt	Analyze this image and respond to: %s	Custom system prompt template
kv_cache_type_k	q4_0	KV cache key quantization type (-ctk)
kv_cache_type_v	q4_0	KV cache value quantization type (-ctv)
idle_timeout	5	Minutes of inactivity before unloading model (0 = disabled)
clipboard_monitor_enabled	false	Enable clipboard history monitor for multi-image analysis
clipboard_history_limit	5	Max cached images in clipboard history (1-20)
clipboard_cache_dir	""	Custom cache directory for clipboard history (default: ~/.go-mcp/vision/clipboard-cache)
model_path	""	Override: exact path to model GGUF file
mmproj_path	""	Override: exact path to mmproj file
llama_server_path	""	Override: exact path to llama-server binary
llama_server_mode	""	Mode: "" (PATH then download), "auto" (download), "custom" (use path)
backend	"local"	Inference engine: "local" (llama-server) or "gemini" (Gemini API)
gemini.api_key	""	Gemini API key from AI Studio
gemini.model	"gemini-3.5-flash"	Gemini model name (e.g. gemini-3.5-flash, gemini-2.5-flash)

Override behavior

When model_path, mmproj_path, or llama_server_path are set to a non-empty path, auto-download is skipped and the specified file is used directly.

llama_server_mode controls binary resolution:

• "" (empty) — search PATH first, download if not found
• "auto" — always download regardless of PATH
• "custom" — use the exact path from llama_server_path

Available Tools

Image Analysis (files, URLs, data URIs)

Tool	Parameters	Description
analyze_image	prompt (required), image (required)	Analyze an image with a custom prompt

Current Clipboard

Tool	Parameters	Description
analyze_clipboard	prompt (required)	Analyze the image currently in the clipboard

Clipboard History (requires clipboard_monitor_enabled: true)

The clipboard monitor runs in the background, polling the clipboard every 500ms. Each new image is detected, deduplicated by hash, and saved to a circular buffer. When the image comes from a file (copied via File Explorer), the original file path is stored — no copy is made, preserving full quality. When the image comes from a raw clipboard source (PrintScreen, browser "Copy Image"), a cached PNG is saved.

History is cleared when the server stops. Cached files are deleted; original files are never touched.

Tool	Parameters	Description
list_clipboard_history	—	List all cached clipboard images with index, timestamp, and source path
analyze_clipboard_image	index (required), prompt (required)	Ask a custom question about a specific cached image by index
analyze_clipboard_images	indices (required), prompt (required)	Analyze multiple cached images at once (comma-separated indices, e.g. "1,2,3")

Multi-image example workflow

User copies 3 images (Ctrl+C three times)
User: "compare the first and third images"

LLM → list_clipboard_history()
LLM  ← #1 — 14:03:15 — C:\Users\...\before.png
        #2 — 14:04:22 — C:\Users\...\middle.png  
        #3 — 14:05:10 — C:\Users\...\after.png

LLM → analyze_clipboard_images(indices="1,3",
         prompt="Image #1 is the BEFORE, image #2 is the AFTER. What changed?")
LLM  ← "The before image shows an empty form. The after image shows..."

Image References

The image parameter accepts:

• HTTP/HTTPS URLs — https://example.com/image.jpg
• Local file paths — C:\Users\...\image.png
• Data URIs — data:image/png;base64,...
• file:/// URIs — file:///C:/image.png

MCP Client Configuration

Kilo Code

Add to your project's .kilocode/mcp.json:

{
  "mcpServers": {
    "vision-mcp": {
      "command": "C:\\path\\to\\vision-mcp.exe"
    }
  }
}

OpenCode

Add to ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "vision-mcp": {
      "type": "local",
      "command": ["C:\\path\\to\\vision-mcp.exe"],
      "enabled": true
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "vision-mcp": {
      "command": "C:\\path\\to\\vision-mcp.exe"
    }
  }
}

Cline / Roo Code / Continue

In .vscode/mcp.json:

{
  "mcpServers": {
    "vision-mcp": {
      "command": "C:\\path\\to\\vision-mcp.exe"
    }
  }
}

PI Agent

In your project's MCP configuration:

{
  "mcpServers": {
    "vision-mcp": {
      "command": "C:\\path\\to\\vision-mcp.exe"
    }
  }
}

Development

# Build
go build -o vision-mcp.exe .

# Run all tests
go test ./internal/...

# Lint
go vet ./...

# Format
go fmt ./...

Changing Models

You can choose from three models during --configure or --manual setup:

Model	Params	Description
unsloth/Qwen3-VL-4B-Instruct-GGUF	4B	VL 4B — good quality and speed (default)
unsloth/Qwen3.5-4B-GGUF	4B	Qwen3.5 4B — proven, good speed
unsloth/Qwen3-VL-2B-Instruct-1M-GGUF	2B	VL 2B 1M context — faster, less VRAM

To change models manually, edit config.json:

{
  "repo_id": "unsloth/Qwen3-VL-4B-Instruct-GGUF",
  "quantization": "Q4_K_M",
  "mmproj": "mmproj-F16.gguf"
}

The server will automatically download the new model files on next start. Make sure the mmproj file is correct for your model architecture.

Benchmarks

See Benchmarks for performance comparison of supported models on a Windows RTX 3080 16GB laptop (1088×552 screenshot).

Setup Wizard Flow

The --configure wizard guides you through these steps:

1. Select Model — choose from available vision models
2. Select Backend — CUDA, CPU, Vulkan, or Metal
3. Select Quantization — quality/size tradeoff based on your hardware
4. Clipboard Monitoring — enable/disable clipboard history
5. Summary — review and confirm
6. Agent Setup — configure MCP for detected agents (Kilo Code, OpenCode, etc.)
7. Download Screen — downloads model files and llama-server with progress bars (skips existing files)

Troubleshooting

Problem	Solution
"llama-server not found"	Run without flags — it auto-downloads. Or place llama-server.exe in PATH or ~/.go-mcp/vision/
"llama-server health check timeout"	Check that the model file exists at the path shown in --status. Try --free and restart.
"CUDA error"	Ensure NVIDIA drivers are installed. Run nvidia-smi to verify.
"Out of memory"	Use a lower quantization (Q3_K_M or Q2_K) in config.json
"Model not downloading"	Check internet connection. Files download from HuggingFace.
Clipboard says "no image found"	Copy an image first (screenshot with Snipping Tool, or copy image from browser)
Clipboard history shows "Clipboard monitor not enabled"	Set clipboard_monitor_enabled: true in config.json, or enable it during --configure or --manual setup
Clipboard history empty after restart	Expected — history is cleared on server shutdown for privacy. Enable monitor and copy images again
Copied file image looks lower quality	When copying from a file, the monitor prefers the original file path (no quality loss). If the image was copied from a browser or screenshot (raw clipboard), a PNG conversion is used — this may alter colors or dimensions slightly
Port 8001 already in use	Run vision-mcp --free to kill existing llama-server

License

MIT