PixLint

Lint, curate, and prepare computer-vision datasets — right from your AI assistant.

PixLint is an MCP server that gives AI assistants — Claude, Cursor, VS Code, and any MCP client — direct, conversational access to a complete computer-vision dataset toolkit: analyze quality, find duplicates and label errors, clean and curate, split, augment, convert formats, and export to every major training framework.

It runs locally over stdio, or self-hosted on the internet over authenticated HTTP.

Why PixLint

Most dataset tooling is either a paid SaaS or a heavy GUI app. PixLint is a single, open-source, self-hostable server an AI agent can drive end to end — and it does things others keep behind paid tiers:

• 🩺 Dataset Doctor — one call runs a full diagnostic and returns a prioritized, executable fix plan.
• Label-error detection — automatically surface images that are probably mislabeled.
• Natural-language query — "find blurry images with a person on the left", answered over your data.
• Weak-slice discovery — find under-represented or low-quality slices to collect or augment next.
• Curation that writes a new dataset — clean / filter / remap, not just report.
• Auto-labeling with a pretrained detector, and one-command Hugging Face publishing.

Features

103 operations — 67 tools, 23 resources, 13 prompts.

Quick Start

1. Install

pip install pixlint

Optional extras add heavier capabilities:

pip install "pixlint[torch]"        # embeddings, auto-labeling, label-error detection
pip install "pixlint[huggingface]"  # Hugging Face export + publishing
pip install "pixlint[all]"          # everything

2. Connect your AI assistant

Claude Desktop — claude_desktop_config.json:

{
  "mcpServers": {
    "pixlint": {
      "command": "pixlint",
      "env": { "CV_DATA_DIR": "/path/to/your/datasets" }
    }
  }
}

Cursor / VS Code — .cursor/mcp.json or .vscode/mcp.json:

{
  "mcpServers": {
    "pixlint": {
      "command": "pixlint",
      "env": { "CV_DATA_DIR": "/path/to/your/datasets" }
    }
  }
}

CV_DATA_DIR is the directory PixLint is allowed to read datasets from.

3. Just ask

"Load my dataset at /data/coco_person, give it a readiness report, then clean it and export for YOLO."

Your assistant calls the right PixLint tools in sequence — diagnose, clean, split, export — and hands back a training-ready dataset.

Security

PixLint touches the filesystem and can be exposed to a network, so protections run on every tool call:

• Paths are confined to your configured data directory (reads and writes).
• Credentials come only from environment variables, never tool inputs.
• Per-call rate limiting, concurrency limits, and audit logging.
• Decompression-bomb protection on image decode.
• Optional bearer-token authentication for the HTTP transport.

See the Security Guide for the full threat model and the recommended production checklist.

Documentation

Runnable scripts live in examples/.

License

PixLint is source-available under the PolyForm Strict License 1.0.0 — see LICENSE. You may use it for permitted (noncommercial) purposes; commercial use, redistribution, or modification requires a separate license from the copyright holder. Contributions are welcome via pull request.

_{mcp-name: io.github.amitsingh-24/pixlint}