civitai-gen

civitai-gen

Unified CLI for Civitai's orchestration Workflow API. All generation types share the same workflow lifecycle: submit, poll, download.

Setup

Requires Node.js 18+. Set CIVITAI_API_KEY either as an environment variable or in a .env file in this skill's directory (next to generate.mjs). Copy .env.example to .env to start. Get a key at: https://civitai.com/user/account

Paths below are relative to this skill's directory. Run the scripts from there (cd into it), or prefix with the install path your runtime uses.

Capabilities

Quick Examples

# Image (defaults to Flux.1, 4 images)
node generate.mjs wait --prompt "A knight at sunset" -o ./out
 

Posting What You Generate

Generate an image (the skill saves it locally), then post the local file in one command:

post-image reads the local file, uploads it, and publishes the post — it prints the public URL. Pull mcp-cli.mjs first if you don't have it (curl -fsSL https://mcp.civitai.com/cli -o mcp-cli.mjs); CIVITAI_API_KEY is required.

Flags: --title (defaults to the filename), --detail, --nsfw <level>, --draft (leave unpublished), --json. For video/audio, or to attach a pre-uploaded image by UUID, call create_post directly (node mcp-cli.mjs schema create_post) — and prefer posting the local file over the generate output's remoteUrls[].url, which is a short-lived signed URL that expires quickly.

Choosing an Engine & Model

Read docs/engines.md to pick the right generator. The key split:

Model discovery uses the Civitai MCP server (hosted at https://mcp.civitai.com/mcp) — call its tools to search checkpoints/LoRAs and get AIR URNs. Don't search inside this skill. A LoRA's base model must match the checkpoint's ecosystem.

If the Civitai MCP isn't connected, add it (browse tools need no API key): claude mcp add --transport http civitai https://mcp.civitai.com/mcp

Can't add MCP config in your runtime? Pull the zero-dep CLI (Node 18+) and call tools over the shell instead: curl -fsSL https://mcp.civitai.com/cli -o mcp-cli.mjs node mcp-cli.mjs call search_models '{"query":"anime portrait","type":"Checkpoint","supportsGeneration":true,"baseModel":"SDXL 1.0"}'

Engine availability/params drift — run node generate.mjs engines for the live list and see https://developer.civitai.com/orchestration/recipes for per-engine parameters.

Domain-Specific Docs

For detailed parameters, read the relevant doc:

• Images: --prompt, --model, --resources, --aspect, --resolution, --source-image (img2img). Run node generate.mjs --help for full flag list.
• Videos: --engine, --duration, --video-aspect, --generate-audio. Run node generate.mjs engines for available engines.
• TTS: Read docs/tts.md — built-in speakers, voice cloning, style instructions.
• Music: Read docs/music.md — ACE Step 1.5, lyrics, duration.
• Transcription: Read docs/transcription.md — ASR, timestamps, language hints.
• Experiments: Read experiment.mjs --help — wildcards, parameter sweeps, naming.

Workflow Lifecycle

All generation types follow the same pattern:

1. Build steps — each job becomes a workflow step with a $type (textToImage, videoGen, textToSpeech, aceStepAudio, transcription)
2. Submit — POST to orchestration API, get workflow ID
3. Poll — check status until terminal state (succeeded/failed/expired)
4. Download — fetch output media (images, videos, audio files)
5. JSON summary — clean JSON to stdout with paths, costs, and metadata

Use cost (whatif) to estimate buzz before spending. Use --quiet for agent-friendly output.

Guidelines

• Default to 4 images per prompt. Use -n 1 only when a single image is needed.
• For video: always check cost first. Video costs 500-2000+ buzz per clip.
• Multiple --prompt flags create concurrent steps in one workflow.
• Use --bulk file.json for large batches.
• Use experiment mode for systematic parameter sweeps.

Developer Guide

See CLAUDE.md for architecture, module layout, and how to add new step types.