Live on mcp.wink.bio · bearer-token beta

Identity for your AI agent.

Wink's face, voice, palm and ID-verification APIs, wrapped as a Model Context Protocol server. Drop it into Claude Desktop, Claude Code or any MCP-capable agent. Point it at an image, get a winktag back. Six lines of config.

Get access →See how it works

Drop any image

HEIC from your iPhone, PNG, JPEG, WebP — any format, any size. The local shim resizes and converts; Wink sees a clean JPEG.

Masked by default

Names return as A****x K****m unless a scoped token explicitly asks for full PII. Winktags are the real identifier; the rest is UX garnish.

Real liveness

Opt-in per call. Wink flags static-photo spoofs and low-confidence matches distinctly, so the agent can ask for a clearer shot or a second factor.

Six lines of config.

Install the local shim, paste a config block into Claude Desktop, start talking. The server is on Vercel; Turso holds device caches and audit logs.

Get a token

Fill the form below. Auto-approved, token shown once on screen.

Install the shim

npm install -g wink-mcp-0.2.0.tgz

Wire it into Claude Desktop

Paste a six-line mcpServers block. Restart Claude. Done.

What your agent can do.

recognize_face(image_path, pii_mode?, liveness?, threshold?)

Single image (HEIC, PNG, JPEG, WebP — any size). Returns verdict (Green/Yellow/Red), confidence, winktag, masked name, and liveness signal. Yellow surfaces 2FA options; Red flags liveness failures.

recognize_faces_in_directory(directory_path, recursive?, max_files?, liveness?)

Batch face recognition. Globs a directory, processes with bounded concurrency, returns one row per file in the same flat shape. Default cap 20 files, hard max 50.

recognize_voice(audio_path, pii_mode?, liveness?)

Single voice sample (m4a, mp3, wav, webm, ogg). Same flat output contract as face — verdict, confidence, winktag, masked name. Liveness runs Wink's active phrase-matching check; a pre-recorded sample goes Red.

recognize_voices_in_directory(directory_path, recursive?, max_files?, liveness?)

Batch voice recognition. Same glob-and-aggregate pattern as the face batch, with bounded concurrency and a per-file summary.

Coming next: verify_id_document, recognize_from_video, complete_consent (cross-merchant unlock).

Get access.

Beta is open. Submit the form and you'll get your bearer token on the next screen, along with a one-click tarball download and the Claude Desktop config snippet.

Token is yours — save it. It's shown once.
No email verification for now — treat it like a beta keycard.
We can revoke individual tokens instantly without affecting anyone else.