Face Blur MCP

Anonymize faces in images — GDPR-grade privacy in one tool call. A hosted MCP server for face detection and blurring: dashcam footage frames, street photography, real-estate listings, event photos, user-generated content, training-data pipelines.

Point any MCP-capable agent (Claude, ChatGPT, Cursor, custom agents) at this server and it can anonymize images as a native tool call — no CV pipeline to build, no model to host.

Tools

blur_faces

Pass an image URL (or base64). Every detected face is anonymized and the result is stored and returned as a URL:

{
  "image_url": "https://example.com/street-photo.jpg",
  "style": "blur",
  "strength": 6
}

Three styles: blur (gaussian — natural look for published imagery), pixelate (mosaic — the classic compliance look), box (solid fill — maximum certainty). Strength 1–10. Face boxes are expanded 15% so hairlines, ears, and jawlines are covered too.

Returns the anonymized image URL, face count, and bounding boxes.

detect_faces

Detection only, no modification: face count, bounding boxes, confidence scores, and a contains_faces flag. Use it for compliance triage ("which of these 10,000 images contain people?") before paying for blurring — it's priced accordingly.

Why this matters

Publishing imagery that contains identifiable faces without consent is a compliance risk under GDPR (and CCPA, and a growing list of state laws). Every mapping company blurs faces; most teams that aren't Google don't have a pipeline for it. This is that pipeline, priced per image.

• Detection: YuNet (OpenCV Zoo) — a modern, widely benchmarked face detector that handles small, rotated, and partially occluded faces. Runs on CPU in ~100ms.
• Privacy by design: input images are processed in memory and never persisted. Output images are stored in your run's key-value store, under your account, not ours.
• Deterministic guardrails: SSRF-protected URL fetching, 20 MB / 100 MP limits, EXIF orientation handled correctly (a classic silent-failure in CV pipelines).

Usage (MCP)

{
  "mcpServers": {
    "face-blur": {
      "url": "https://<actor-standby-url>/mcp"
    }
  }
}

Pricing

Pay per image — no subscription:

Event	Tool	Price
face-blur	blur_faces (and batch runs)	$0.02
face-detect	detect_faces	$0.005

Anonymization SaaS platforms charge €0.01–0.06/image with annual contracts and minimums. This is contract-free at the same unit price.

Batch mode

Run the Actor directly with { "imageUrls": [...], "style": "blur" } to anonymize a whole folder of URLs into the dataset — one output URL per image.

Limits

• JPEG/PNG/WebP in; JPEG or PNG out. Max 20 MB / 100 megapixels per image.
• Images only (video frames work — extract frames upstream; native video is on the roadmap).
• Detection is probabilistic. For regulated publication, spot-check output — no detector catches 100% of faces 100% of the time.