jsontoimg

MCP Server for Image Generation: The Render Loop in 11 Tools

Published October 15, 2026 by Balaji

Uncategorized

An MCP server for image generation lets an agent in Cursor, Claude Desktop, or VS Code render images without hand-written HTTP clients. jsontoimg’s server exposes eleven tools shaped around one loop: find a template, read its schema, render, and fetch the result. It is not a text-to-image model; the agent fills a template you designed.

Connect

Create an API key, open Integrations, then MCP, in the dashboard for copy-paste config, and point your client at the HTTP endpoint at /mcp. The endpoint accepts an API key only; sessions are not accepted. The Model Context Protocol specification describes the transport, and Cursor documents client setup in its MCP guide.

For background on how clients discover and call tools, read the MCP introduction.

The eleven tools

Tool Purpose
list_templates A page of templates (name and description); pass offset while hasMore is true
get_template_schema Named layers and attributes for a template
create_render Queue a render
wait_for_render Long-poll until done or failed
get_render Job status
get_render_asset_url Signed download URL that works in <img>
create_batch Queue up to 50 jobs
get_batch Batch status and child jobs
sign_image_url Mint a signed CMS or Open Graph URL
sign_image_urls Mint up to 50 signed URLs; does not queue renders
get_signed_image Fetch a signed image and return its bytes

The agent flow

  1. list_templates, then pick the template that matches the prompt.
  2. get_template_schema to learn the layer names.
  3. create_render with layers, then wait_for_render.
  4. Return get_render_asset_url to the user.

For <img> or og:image links, call sign_image_url instead. That path mints a URL without queuing a render.

Why a small surface helps

An agent chooses tools by their names and descriptions, and a broad suite adds ways to choose wrong. This server does the render loop and nothing else: no social publishing, no video, no account management. Give each template a clear description and name its layers like fields, headline and price, because the schema is all the agent sees.

What to watch

  • Credits. Each render consumes the same credits as any API call. Put a spending cap on any agent you leave unattended.
  • Template quality. A template with layers named text1 and text2 produces bad renders from a good agent.
  • Scope. The agent can only fill layers. It cannot redesign the template.

The agent-side overview is on the agents page, and the other trigger types are in the Image Automation API guide.

FAQ

What is an MCP server for image generation?

A server that exposes image rendering as tools an AI agent can call. jsontoimg’s has eleven tools covering templates, renders, batches, and signed URLs.

Is the jsontoimg MCP server an AI image generator?

No. The agent fills named layers on a template you designed. It does not invent images from a prompt.

How do I authenticate to the MCP server?

Use an API key against the HTTP endpoint at /mcp. Sessions are not accepted. Copy-paste config is in the dashboard under Integrations, then MCP.

Can an agent render Open Graph images?

Yes. It calls sign_image_url, or sign_image_urls for up to 50, to mint signed URLs without queuing renders.

Related Resources