Local Image Search

Give your AI coding agent the ability to search through all your local images. Privacy-first, 100% local MCP server for macOS. Uses MLX CLIP for embeddings, Daft for batch processing, and Lance for vector storage.

https://github.com/user-attachments/assets/41e167f0-bb73-4310-8c1c-4be07af21cc1

Features

• 100% local - Images and embeddings never leave your machine
• MCP Server - Works with Claude Code and Claude Desktop
• Natural language search - Find images by describing them
• Fast - 260+ images/second on Apple Silicon via MLX

Requirements

• macOS with Apple Silicon (M1/M2/M3/M4)
• uv (for uvx command)

Quick Start

Claude Code

Option 1: CLI

claude mcp add local-image-search -- uvx local-image-search

Option 2: Manual - add to ~/.claude.json:

{
  "mcpServers": {
    "local-image-search": {
      "command": "uvx",
      "args": ["local-image-search"]
    }
  }
}

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "local-image-search": {
      "command": "uvx",
      "args": ["local-image-search"]
    }
  }
}

Restart Claude after setup. The first run downloads the model (~600MB) and embeds your images, which may take a few minutes. After that, it only processes new or changed files. By default, it scans your home directory (~) and skips common system folders. See Configuration Logic for details.

Custom Configuration

Scan a specific folder:

{
  "args": ["local-image-search", "~/Pictures"]
}

Custom excludes:

{
  "args": ["local-image-search"],
  "env": {
    "EXCLUDE_DIRS": "Downloads,Desktop,Movies"
  }
}

Faster refresh:

{
  "env": {
    "REFRESH_INTERVAL": "30"
  }
}

Configuration Logic

Options	Root	Excludes
None	~ (home)	Default excludes
Root only	Custom root	None
Excludes only	~ (home)	Custom excludes
Root + Excludes	Custom root	Custom excludes

Default excludes: Library, .Trash, .cache, Cache, node_modules, .git, .venv, venv

MCP Tools

• search_images(query, limit) - Search for images matching a text description
• get_status() - Check if the service is ready (model loaded, embeddings synced)

Development Setup

# Clone the repo
git clone https://github.com/Eventual-Inc/local-image-search.git
cd local-image-search

# Install dependencies
uv sync

# Download and convert CLIP model (~600MB, first time only)
cd clip && uv run python convert.py && cd ..

CLI Usage

Embed images from a directory

uv run python embed.py ~/Pictures           # embed all images
uv run python embed.py ~/Pictures --dry-run # count and estimate time
uv run python embed.py . --no-recursive     # current dir only

Embeddings are cached in embeddings.lance/. Re-running skips unchanged files.

Supported formats

Format	Extensions	Tested
JPEG	.jpg, .jpeg	Created and embedded
PNG	.png	Created and embedded
GIF	.gif	Created and embedded
WebP	.webp	Created and embedded
BMP	.bmp	Created and embedded
TIFF	.tiff, .tif	Created and embedded
HEIC/HEIF	.heic, .heif	Real iPhone photo + converted PNG

Corrupted or unreadable images get zero vectors (won't match searches).

Search

Start the server (loads model once):

uv run python server.py

Search via CLI:

uv run python search.py "sunset"           # list results
uv run python search.py "people" -n 10     # show 10 results

Or via API:

curl -X POST http://127.0.0.1:8000/search \
  -H "Content-Type: application/json" \
  -d '{"query": "yellow mouse", "limit": 5}'

Demo scripts

uv run python simple_image_search.py  # basic in-memory search (2 images)
uv run python daft_image_search.py    # batch processing demo

Project Structure

local-image-search/
├── clip/                    # MLX CLIP implementation (from ml-explore/mlx-examples)
│   ├── model.py             # CLIP model architecture
│   ├── clip.py              # Model loading and inference
│   ├── convert.py           # HuggingFace to MLX converter
│   ├── image_processor.py   # Image preprocessing
│   ├── tokenizer.py         # Text tokenization
│   ├── mlx_model/           # Converted model weights (generated)
│   └── LICENSE              # MIT License (Apple Inc.)
├── data/
│   └── pokemon/             # Pokemon artwork (1025 images)
├── embeddings.lance/        # Lance DB storage (generated)
├── mcp_server.py            # MCP server entry point
├── server.py                # FastAPI server for local API
├── search.py                # CLI search tool
├── core.py                  # Shared utilities (EmbedImages, find_images, etc.)
├── embed.py                 # CLI tool to sync embeddings from a directory
├── test_embed.py            # Tests for embed.py
├── simple_image_search.py   # Basic in-memory search demo
├── daft_image_search.py     # Daft-based batch processing demo
├── benchmark.py             # Benchmark script
├── plot_benchmark.py        # Generate benchmark plot
├── benchmark_results.csv    # Raw benchmark data (10 runs)
├── benchmark_plot.png       # Benchmark visualization
├── pyproject.toml           # Project dependencies
└── uv.lock                  # Dependency lockfile

Benchmarks

Embedding time for the Pokemon dataset (1025 images) on M4 Max, averaged over 10 runs.

Run benchmarks yourself:

uv run python benchmark.py      # Run one iteration, appends to CSV
uv run python benchmark.py 100  # Benchmark with specific number of images
uv run python plot_benchmark.py # Generate plot from CSV

Real-world performance (M4 Max, home directory)

Metric	Value
Images found	11,843
Scan time	~26s
Embed time	~39s
Total time	~65s
Embed speed	260 img/s
Re-run (cached)	~31s (scan only)

Data Attribution

Pokemon Artwork

• Source: PokeAPI/sprites
• License: Repository is CC0 1.0 Universal
• Copyright: All Pokemon images are Copyright The Pokemon Company

CLIP Implementation

• Source: ml-explore/mlx-examples
• License: MIT License (Apple Inc.)