Cascade Img

Generate Midjourney images by conversation instead of by hand. You describe what you want; your AI assistant composes the prompt, fires it, inspects the grid with vision, crops the best quadrant, cleans it up, and logs what worked.

You: "I need a flat-design mountain icon, centered, simple shapes, transparent background"

Agent: reads prompt log → composes prompt from parts → fires imagine →
       waits → inspects 2x2 grid with vision → picks best quadrant →
       crops it → removes background → saves → logs what worked

cascade-img is an MCP server with 23 tools that plugs into Claude, Cursor, Codex, or anything that speaks MCP. Midjourney is the first backend; Flux, DALL-E, and Imagen are on the roadmap. There's also a CLI.

Not a programmer? Open an AI assistant that can run commands (Claude Code, Cursor, or Cline), point it at this repo, and say: "Read RUNBOOK.md and set up cascade-img on this machine, then let me make images by describing them to you." It does the technical parts. You just need a Midjourney subscription and to copy a few values from Discord.

Quick Start

You need: a paid Midjourney subscription, a Discord account with the MJ bot in a channel, and Python 3.12 or newer.

pip install cascade-img

Or from source:

git clone https://github.com/laffeyp/cascade-img
cd cascade-img/packages/python
pip install -e .

This puts three commands for operating cascade-img on your PATH: cascade-mj-bridge (the daemon), cascade-mcp (the MCP server), and cascade-mj (the CLI). Installing also adds a fourth command, cascade-trace-check — a diagnostics validator (not part of the generation loop) that replays a recorded event log and checks it against the vocabulary's declared event ordering and timing rules.

Configure — you need four values from the Discord desktop app (channel ID, server ID, imagine version, and your user token). Takes about five minutes. RUNBOOK.md walks through each one step by step.

cp "$(python -c 'import cascade_img, pathlib; print(pathlib.Path(cascade_img.__path__[0]) / ".env.example")')" .env
# Fill in the four values per RUNBOOK.md, then validate:
cascade-mj-bridge --check-env --pretty

Start the daemon in one terminal, then connect from another:

cascade-mj-bridge          # leave running — holds the Discord connection

Connect your AI assistant — add to your MCP config (Claude Desktop, Cursor, Cline):

{
  "mcpServers": {
    "cascade-img": {
      "command": "cascade-mcp"
    }
  }
}

Or point your assistant at this repo and ask it to read AGENTS.md — it'll wire everything up.

Or use the CLI:

echo '{
  "mountain-icon": {
    "subject": "a flat-design icon of a mountain, centered, simple shapes",
    "aspect_ratio": "1:1"
  }
}' > assets.json

cascade-mj mountain-icon --registry assets.json --upscale all --pretty

The 23 Tools

Every call returns {ok, result} or {ok: false, error: {code, remediation}}. Branch on the stable code, not the message. Full tool reference in AGENTS.md.

How This Differs

Other open-source Midjourney tools focus on the generation step — fire the prompt, hand back the image. cascade-img does the work around that:

• Vision-based self-curation — the agent inspects its own output and picks the best quadrant
• Structured prompt composition — prompts built from parts (subject, style, identity, constraints), not raw strings
• Working memory — append-only log persists across sessions; each run reads what came before
• Curation pipeline — crop grids, remove backgrounds, build sprite sheets, promote winners
• MCP-native — 23 tools that plug into Claude, Cursor, Codex, or anything that speaks MCP
• Pluggable backends — Midjourney now, Flux/DALL-E/Imagen on the roadmap

How It Works

One daemon, two stateless clients, all over local HTTP:

• cascade-mj-bridge — the daemon. Only process that talks to Discord. Holds the live connection and tracks in-flight jobs. Must stay running.
• cascade-mcp — the MCP server. Stdio by default (Claude Desktop / Cursor / Cline); --http <port> for HTTP. Stateless — start and stop freely.
• cascade-mj — the CLI. Takes an asset ID and a registry, composes the prompt, fires, waits, writes to the log.

Prompts are composed from structured parts, not written as raw strings:

from cascade_img.prompt.composer import PromptComposer, Subject, StyleStack, IdentityStack

prompt = PromptComposer().compose(
    Subject(
        text="a flat-design icon of a mountain",
        constraints=["centered", "simple shapes", "transparent background"],
    ),
    # Both optional. moodboard is a Midjourney personalization code;
    # sref/oref are reference-image URLs you'd set up in MJ first.
    style=StyleStack(moodboard="abc123def", sref="https://cdn.example.com/style.png"),
    identity=IdentityStack(oref="https://cdn.example.com/ref.png", ow=1000),
    aspect_ratio="1:1",
    version="7",
)

All three entry points emit structured JSON and follow the same {ok, result | error: {code, remediation}} envelope. Every failure carries a stable error code (e.g. DISCORD_401, MJ_UUID_MISSING, UPSCALE_BUTTON_FAILED) with a machine-readable remediation — so a caller branches on the code, not the message. The full catalog of log events and error codes is in vocabulary/0.1.json, and a trace checker (cascade-trace-check) enforces event ordering over recorded runs.

Documentation

Roadmap