Skip to content
~/skunk $ cat imagen-mcp.md

Imagen 4 MCP Server

Imagen 4 + Claude, actually talking to each other.

STATUS Active (v2)TypeScriptNodeGoogle Imagen/Gemini APIMCP

Private repo — publishable.

GOAL

Expose Google's Imagen 4 and Veo video generation directly inside Claude as a Claude Code plugin.

NEEDS

Stop copy-pasting between a chat and a web UI — "generate this" and it happens, in-session.

PURPOSE

Ship production AI tooling at the bleeding edge, native to the Claude Code plugin system.

This is an MCP server, packaged as a Claude Code plugin, that puts Google's Imagen 4 and Veo directly into a Claude Code session. Imagen 4 ships in three tiers — Fast, Standard, Ultra — trading cost and latency for fidelity, and Veo handles video generation from the same plugin. No browser tab, no downloading a file and re-uploading it into a chat: ask for an image, get it back in-session, iterate immediately.

The point wasn't just "call an API" — it was making image generation a native tool a coding agent reaches for mid-task, the same way it reaches for a file edit or a shell command. That meant handling auth, cost tracking across tiers, and sane defaults so a plugin install is the only setup step.

The comparison below is the honest way to show it: one reference photo, a black 1956 Plymouth Belvedere shot in a workshop, run through all three Imagen 4 tiers back to back. Fast lands around $0.02 a generation, Standard around $0.04, Ultra around $0.06 — and the jump in fidelity from tier to tier is exactly what you'd hope to see for the price.

~/skunk $ ls ./gallery
Reference photo — a black 1956 Plymouth Belvedere in a workshop
The reference photo the models had to recreate
Imagen 4 Fast recreation of the Belvedere
Imagen 4 Fast — ~$0.02
Imagen 4 Standard recreation of the Belvedere
Imagen 4 Standard — ~$0.04
Imagen 4 Ultra recreation of the Belvedere
Imagen 4 Ultra — ~$0.06

← Back to projects