All articles

Guides

Face recognition MCP server: find every photo of a person from Claude, Cursor or VS Code

Connect one URL, index your photos, and ask your assistant which photos a person appears in.

By SightRadar EngineeringUpdated 6 min read Markdown

TL;DR

The SightRadar MCP server gives an AI assistant 39 face recognition tools over your own photo collections. Add https://mcp.sightradar.com/mcp to Claude, ChatGPT, Cursor or VS Code, sign in, and ask in plain language: index these photos, then find every photo this selfie appears in. Calls that process a photo spend credits at the normal API price, per photo and never per face; reading and listing are free.

The Model Context Protocol (MCP) lets an assistant call tools for you. With a face recognition server connected, jobs that used to need a script become one sentence: build a collection from a folder of event photos, find one guest in all of them, check whether two faces are the same person.

Who already runs this face search

sellphotos.online is a marketplace where event and sports photographers upload a shoot and guests find their own photos with a selfie. Their search runs on the same SightRadar API that the MCP server wraps. They call the API from their own code, not through an assistant, but the operations are the same ones you get as tools: index photos into a collection, then search it with a selfie.

They moved from AWS Rekognition in about two days, and by their figures the monthly bill went from about $3,500 to about $1,550. The full story is in the sellphotos.online case study.

Connect it in one step

The hosted server needs nothing installed. Add the URL to your client. The first time the assistant uses it, your browser opens sightradar.com to sign in and you choose what the assistant may do.

  • Claude (claude.ai or Claude Desktop): Settings, Connectors, Add custom connector, paste https://mcp.sightradar.com/mcp.
  • ChatGPT: Settings, Apps & Connectors, Create, paste the URL and pick OAuth.
  • Cursor: add the server to ~/.cursor/mcp.json.
  • VS Code: add it to .vscode/mcp.json under servers, with type set to http.
claude mcp add --transport http sightradar https://mcp.sightradar.com/mcp
Claude Code
{
  "mcpServers": {
    "sightradar": { "url": "https://mcp.sightradar.com/mcp" }
  }
}
~/.cursor/mcp.json

If your photos are on your computer, use the local server instead. It runs the same tools over stdio and can read image files with file_path. It needs Node.js 20 or newer and an API key from the console.

claude mcp add sightradar --env SIGHTRADAR_API_KEY=frs_your_key -- npx -y sightradar-mcp
Local server in Claude Code

Find a person's photos in three prompts

  1. Create and fill a collection. "Create a collection called wedding-2026-10 and index these photo URLs." For more than a handful of photos the assistant uses submit_batch, which takes up to 1,000 photo URLs in one call and runs in the background.
  2. Search with a selfie. "Which photos in wedding-2026-10 is the person in this selfie in?" The assistant calls search_faces and returns the matching photos with a similarity score for each.
  3. Search again without the image. If you registered the selfie with register_selfie, search_by_point_id repeats the search later without sending the picture again.

Tip: The hosted server takes images as a public URL or as base64 (up to 10 MB). For large files, prefer a URL.

What each call costs

The tools spend credits at the same prices as the REST API, and every billable tool states its price in its description. Billing is per photo processed, never per face, so a group photo with twelve faces costs the same as a portrait.

ToolWhat it doesPrice
index_photoStore one photo's faces in a collection$0.00093 per photo
submit_batchIndex or match up to 1,000 photo URLs in the background$0.00062 per photo
search_facesFind every photo that contains the person in a selfie$0.00093 per search
search_by_point_idRepeat a registered selfie's search$0.00062 per search
compare_facesCheck whether two faces are the same person$0.00093 per image
detect_facesFaces, bounding boxes and quality in one image$0.00093 per image
Prices from the SightRadar rate card. Creating collections and webhooks is free.

Limit what the assistant can do

The 39 tools are split into five permission groups: read, search, write, delete and account. During sign-in you tick the groups this connection gets. A read-only connection cannot spend credits or change data, and account always starts unticked.

  • Each connection gets its own API key, named MCP: <app name>. Revoke it in the console and the assistant is cut off.
  • Deletes are soft by default, with a 7-day restore window.
  • Tools carry MCP annotations, so clients that support them ask you before a destructive call.

Note: A face is personal data. Index photos you have the right to process, and register a person's selfie only with their consent, as the Acceptable Use Policy requires.

Setup for every client, the full tool list and the permission groups are in the docs.

Read the MCP docs

Frequently asked questions

Is there an MCP server for face recognition?

Yes. SightRadar runs a hosted MCP server at https://mcp.sightradar.com/mcp and publishes the same server as the sightradar-mcp npm package. It gives an assistant tools to index photos, search a collection with a selfie, compare two faces and detect faces.

Which AI assistants does it work with?

Any MCP client. The docs cover Claude, ChatGPT, Claude Code, Cursor and VS Code. Clients that only run local servers can use the sightradar-mcp npm package.

Does the assistant see my API key?

No. With browser sign-in, SightRadar creates a key for the connection and hands it to the MCP server directly. The key never passes through the browser or the assistant.

Is the MCP server free?

The server is open source under the MIT licence and costs nothing to connect. The tools that process photos spend credits at the normal API price, per photo and never per face.

Can I use it on photos on my computer?

Yes, with the local server. It reads image files from folders you allow. The hosted server takes public image URLs or base64 data.

Keep reading