For vibe coders: agents and the manifest

You want to describe what you want and have an agent do it. Image Forge is built for that: the manifest is a plain CSV, and there is an MCP server that exposes the whole pipeline as eight tools.

Wire it up in one command 

claude mcp add image-forge node scripts/mcp-server.js

Or drop this in your project’s .mcp.json:

{
  "mcpServers": {
    "image-forge": {
      "command": "node",
      "args": ["scripts/mcp-server.js", "--csv", "marketplace-images.csv", "--out", "generated-images"]
    }
  }
}

Works with Claude Code, Cursor, Cline, Windsurf, n8n, LangChain — anything that speaks MCP over stdio.

Then just talk to it:

“Look at the manifest and add eight tavern interiors in different regional styles, then forge the pending ones.”

The eight tools

Tool What it does
forge_status How many pending, done, failed — and which engine is wired up
forge_list List rows, optionally filtered by status
forge_add_row Add a picture idea
forge_generate_pending Generate everything pending, with an optional limit
forge_generate_one Generate one row by filename
forge_retry_failed Put failed rows back to pending
forge_fix_retired Move rows off models the provider switched off
forge_models Every model with price per image, delayed price, free allowance

forge_generate_* write real files into images/, vectors/, lottie/, sheets/ and gifs/ under --out, and update the CSV. They are not simulations.

Giving the agent engines

With no configuration at all it uses OVHcloud SDXL, which needs no key. Two pictures a minute, always square.

To give it the same engines the app has, point it at a backup file from the app — either Settings → Back up to a file or Settings → Advanced → Backup:

node scripts/mcp-server.js --settings ./image-forge-settings-2026-09-13.json

Or use environment variables:

Variable Meaning
CLOUDFLARE_ACCOUNT_ID · CLOUDFLARE_API_TOKEN Cloudflare Workers AI, free
POLLINATIONS_TOKEN Pollinations, free
GEMINI_API_KEY / GEMINI_API_KEYS Google keys, comma-separated, rotated on 429
GEMINI_IMAGE_MODEL Which Google model, e.g. nano-banana-2-lite
OPENAI_API_KEY / OPENAI_API_KEYS Keys for any OpenAI-compatible endpoint
OPENAI_BASE_URL · OPENAI_IMAGE_MODEL Point at Together, OpenRouter, a local server
FORGE_PROVIDER Force ovh / cloudflare / pollinations / gemini / openai

Which engine it uses when a row names none:

  1. FORGE_PROVIDER, if set.
  2. Otherwise the engine that was selected when the backup was made (the practice forge is skipped — an agent writing real files should not produce practice drawings).
  3. Otherwise the first one that is set up: Cloudflare, Pollinations, Google, OpenAI-compatible.
  4. Otherwise OVHcloud, which needs nothing.

A row’s own model column beats all of it, exactly as in the app.

Things worth knowing before you let it run

Give it a free engine. An agent in a loop is exactly how you discover what your paid balance was. The app’s paid confirmation dialog does not exist over MCP — there is no human there to confirm — so the server has no spending guard. Keep paid keys out of its environment unless you are watching.

Cap the batch. forge_generate_pending takes a limit. Use it.

OVHcloud is slow on purpose. Two a minute. A limit of 10 takes about five minutes.

Weak models cannot spell. The server suppresses text instructions on models that produce gibberish, the same way the app does. Ask a Nano Banana model if you need real words.

Filenames are validated. forge_add_row rejects path separators outright. An agent cannot write outside the output folder by choosing a clever filename.

The CSV is the real API

If MCP is not your thing, skip it. The manifest is RFC 4180 CSV, and any tool that can read and write it participates fully:

import csv

with open("marketplace-images.csv", newline="", encoding="utf-8") as f:
    rows = list(csv.DictReader(f))

pending = [r for r in rows if r["status"] == "pending"]

Missing columns are forgiven on import, so you can hand it a two-column CSV with just filename and prompt.

Wiring into other stacks

CONNECT-AGENTS.md covers Claude Code, n8n, LangGraph and Hermes.

For LangGraph: treat forge_generate_pending as a node, forge_status as your loop condition, and let the CSV be the state that survives between runs.

Where to go next