AI Video Producer
April 26, 2026 · View on GitHub
A Claude Code plugin that drives the end-to-end lifecycle of a single AI-generated video project: creative brief → model selection → character sheets → script → storyboard → generation pipelines → clip assembly → post-production → final export.
This is a state-aware lifecycle plugin — it operates inside a per-project workspace it scaffolds itself. Distinct from utility-belt plugins that fire stateless one-off operations on media files.
What's in the box
Commands
/install-video-workspace— scaffold the per-project folder structure + project CLAUDE.md into the current directory./setup-credentials— one-time setup for fal.ai, Replicate, WaveSpeed, and MiniMax API keys, plus runner toolchain bootstrap (npm + pip). Writes~/.config/ai-video-producer/.envand wires it into your shell profile so the bundled MCP servers always have keys./onboard— capture the creative brief and tools-and-models selection./define-character— add a recurring character/subject sheet./draft-script— produce a script draft from the brief./storyboard— break the locked script into a numbered shot list with per-shot briefs./run-pipeline <name> <shot#>— execute a predefined pipeline for a shot./log-take <path>— record a generated artefact's prompt, model, parameters, and cost./promote-take <path>— copy a raw take intoclips/selected/with rationale./assemble— propose an ffmpeg edit plan to concatenate clips into a rough cut./export-final— encode the locked draft to spec, embed subtitles, write tooutput/final/.
Agents
Spawned via the Task tool when a job warrants a dedicated subagent.
script-writer— drafts and revises scripts from the brief; iterates with the user; promotes a chosen draft toscripts/final/script.md.pipeline-scaffolder— designs a new pipeline as a structuredSPEC.md(planning only; no code).pipeline-builder— converts an approved SPEC into runnable per-stage runners.pipeline-repo-setup— provisions a private GitHub repo for versioning reusable pipelines across projects.concatenator— joins clip elements into a single render. GPU-aware (NVENC / AMF / VAAPI / QSV / CPU). Picks fast-path (concat-demuxer) vs safe-path (re-encode).normalizer— equalises loudness (EBU R128, two-pass), conforms framerate / pixel format / colour tag, writes toclips/normalised/.aspect-converter— aspect ratio conversion (16:9 ↔ 9:16, blur-fill / letterbox / smart-crop) and resolution upscaling. GPU-aware. Defers to AI-upscale skills when ≥2× uplift is needed.
Skills
Pipeline recipes:
text-to-image-to-video— Flux/SDXL/Imagen → Kling/Runway/Hailuo.voice-to-lip-sync— voice generation → lip-sync.text-to-video-direct— single-step text-to-video.upscale-and-interpolate— resolution + frame-rate uplift.video-workspace-conventions— reference for the workspace layout.
Production utilities:
script-to-tts— reformat a finished script for a specific TTS model (ElevenLabs, OpenAI, Google, Azure, Hume, Chatterbox), with model-appropriate SSML / pacing / chunking.model-researcher— research current fal/Replicate (and direct-provider) offerings for a workload; recommend a best-fit model with rationale; updatebrief/tools-and-models.md.budget-estimator— project API cost (low / typical / high) for a pipeline or whole script before any generation runs.comfyui-import— turn a ComfyUI workflow JSON into a pipeline SPEC + stage runners (with custom-node + checkpoint manifest).comfyui-export— turn a pipeline SPEC into a ComfyUI API-format workflow JSON for local execution; flag stages that have no clean local equivalent.
Bundled MCP servers
The plugin ships with three MCP servers preconfigured (see .mcp.json):
- fal-ai — hosted at
https://mcp.fal.ai/mcp(HTTP transport, requiresFAL_KEY). Catalogue browsing only; actual generation goes throughrunners/fal_run.mjs. - replicate — local stdio via
npx -y replicate-mcp(requiresREPLICATE_API_TOKEN). Used for both browsing and execution. - minimax — local stdio via
uvx minimax-mcp(requiresMINIMAX_API_KEY+MINIMAX_API_HOST). Providesgenerate_video(Hailuo),text_to_image,text_to_audio,voice_clone,voice_design,music_generation.
Credentials live in ~/.config/ai-video-producer/.env. The /setup-credentials command creates that file and adds a loader line to your shell profile so every shell that launches Claude Code has the values in its environment. See .env.example for the full template.
SDK runners
For execution-heavy providers, the plugin ships SDK wrappers under runners/ instead of relying on MCP catalogue servers:
- fal-js (
runners/fal_run.mjs,@fal-ai/client) — queues jobs, polls, downloads results. RequiresFAL_KEY. - WaveSpeed Python (
runners/wavespeed_run.py,wavespeed) — same shape, forwavespeed-ai/*models. RequiresWAVESPEED_API_KEY.
Both runners share a uniform CLI contract — see runners/README.md. They print a JSON envelope to stdout ({model, output_files, data}) and stream progress to stderr, so any pipeline skill can shell out and parse the result.
/setup-credentials installs both toolchains (npm install for fal-js, a venv + pip install -r requirements.txt for wavespeed).
Typical workflow
mkdir my-video-project && cd my-video-project
claude
> /setup-credentials # once per machine
> /install-video-workspace
> /onboard
> /define-character
> /draft-script # or spawn the script-writer agent
> /storyboard
> /run-pipeline text-to-image-to-video 01
> /promote-take clips/raw/01-establishing-v3.mp4
... repeat for each shot ...
> /assemble # rough cut
... or spawn `normalizer` → `concatenator` → `aspect-converter` for finer control ...
> /export-final
Recommended companion plugins
- video-editing — ffmpeg utilities (transcode, watermark, merge) for ad-hoc clip work.
- video-production — ComfyUI workspace management and cover-art generation.
Required external tools
- ffmpeg (system binary) — used by
concatenator,normalizer,aspect-converter,/assemble,/export-final. - gh (authenticated) — used by
pipeline-repo-setup. - GPU drivers — NVIDIA (CUDA + NVENC), AMD (ROCm + AMF/VAAPI), or Intel (QSV/VAAPI). Optional; CPU encoding is the fallback.
Installation
claude plugins install ai-video-producer@danielrosehill
License
MIT