DramaClaw: A Source-Available AI Drama Production Line with an Infinite Canvas, a Dual-Track Asset Library and an MCP Agent

What it is

DramaClaw is a source-available production line for AI-generated short drama. The team behind it says this is the same industrialized pipeline it runs every day, not a demo or a cut-down edition. It covers the whole chain: manuscript or screenplay in, structured episodes, scripts, storyboards, first frames, voice-over and a finished film out. It ships with an infinite node canvas and a built-in agent, and you host it yourself.

The license is Elastic License 2.0. You may run it, modify it and sell what you make with it. The project asks for a small "Powered by DramaClaw" mark in your UI. The one thing still closed is reselling it as a hosted SaaS for other people.

Core architecture: two tracks, one asset library

The central design choice is a dual track. The first track is XiaHua, a node-based infinite canvas. The README lists 18 node types: upload, image generation and edit, storyboard generation, script, beat context, video, video compose, video story, audio, style, skill, group, text annotation, a 360-degree panorama viewer, a 3D world node, export and more. Every node keeps its own generation history, and nodes connect freely. For large projects the canvas adds tabs, an element outline, a minimap, viewport bookmarks, snap-align, level-of-detail rendering, revision history with restore, and per-canvas locking.

The second track is the series pipeline, called XiaJi. It runs in four stages: ingest, plan, produce, deliver. Ingest builds episodes, characters and scenes directly from a manuscript or a Fountain screenplay, with no knowledge graph or embeddings required. A legacy path based on the Cognee story graph remains for older projects. Planning handles chapter segmentation, beat planning and multi-episode arcs. Script generation offers adaptive, literal and staged modes, each with a review-and-repair loop. Production creates storyboards and first frames, plus emotion-aware voice-over. Delivery assembles episodes and exports video, subtitles and the full asset pack.

Both tracks share one asset library, called XiaTang. You explore freely on the canvas, preview the impact, then promote a node or a batch of nodes into the library or an episode. You can also project the series back onto a canvas from presets. Most tools offer only a free canvas or only a rigid wizard. Running both over one library is the project's main structural idea.

How it works inside

Each pipeline step is an independent asynchronous task with its own interface. You can run steps in order, skip some, resume from any checkpoint, or plug in your own orchestrator. The Task Center shows status, progress and logs, supports cancel, retry and resume, and validates prerequisites before it queues a job. The infrastructure is deliberately small. There is no Postgres, Redis, Celery or Ray. Tasks run in process, and state lives in SQLite and plain files. Because of that, 2 vCPU and 4 GB of RAM is the recommended minimum, and the standard pipeline needs no GPU. Only the optional world extra, which does voxel and panorama-to-3D work, needs a GPU and a CUDA image.

No model runs on your machine. All inference goes through an OpenAI-compatible gateway, and you choose one of three modes in Settings. Official mode uses the hosted RelayClaw service with a DC key and needs no mapping. Custom mode initializes the bundled dramaclaw-gateway, which you fill with your own provider channels. Hybrid mode uses the official models for the main pipeline and adds extra channels, for example a local ComfyUI video workflow.

The bundled gateway is a fork of New API. It implements the project's DC-Media contract, which expresses media roles, reference files and first and last frames, and then translates each request into the provider's native API. Adapters listed today cover ComfyUI, MiniMax and Hailuo, VolcEngine Doubao and Seedance, fal.ai, Alibaba, Kling, Jimeng, Vertex AI, Gemini, OpenAI and Sora, and Suno. In official mode the README names gpt-image and nano-banana for images, the Seedance 1.0, 1.5 and 2.0 series for video, and IndexTTS2 for voice-over. This contract is the layer that keeps the project model-neutral.

Spatial consistency: Director World

Consistency across shots is the hardest problem in generated drama. Director World addresses it with a virtual set. It turns an image into a 3D Gaussian Splat scene and adds scene-360 panoramas as canvas nodes.

The set locks spatial structure, character blocking and camera placement. You frame the shot in the 3D viewer, capture it, and use the capture as the background for the next generation. The same location then stays stable from shot to shot. This borrows a real film workflow, build the set first and shoot second, and moves it into a generative pipeline.

The agent and the ecosystem

The built-in agent is called Xia Director. Today it knows your project, checks progress, advances script and shot tasks, audits whether deliverables are complete and suggests next steps. A local MCP server exposes DramaClaw to Claude Code, Codex and any other MCP client. It listens on loopback only and needs an explicit trust flag.

The README shows one prompt to Xia Director producing a 17-node, 32-edge workflow: brief, story direction, character and set references, five first frames, five shot videos, voice-over and score, final cut. Read this carefully. The agent building and running canvas graphs sits under "In Development", and each canvas command needs approval in the browser first. Other items in development include a catalog of 11 workflow skills and 60 recipes, an agent-kit for Hermes, OpenClaw, Claude Code and Codex, a browser-based 3D previz stage, and branching interactive stories and ads.

Performance, cost and what is not proven

The repository publishes no reproducible benchmark. It gives no per-episode time, no cost per minute of video and no quality scores. Cost is set mostly by the image, video and voice models behind your gateway, while the local footprint is small.

The README also includes a comparison table against seven masked competitors, with checkmarks for storyboards, asset libraries, canvas, agents and more. The project authors scored it from public product documents. Treat it as positioning, not independent evaluation.

Limits and risks

First, Elastic 2.0 is source-available, not open source by the OSI definition. Hosted SaaS is barred, and the attribution request applies to the standalone version. Second, output quality depends on upstream video models.

Character identity, lip sync and long-form coherence remain open problems for every generative approach. Third, the most ambitious part, an agent that builds canvas graphs by itself, is not in a release yet. Fourth, official mode sends traffic through relayclaw.cdnfg.com, so privacy and uptime need your own review, while custom mode makes you maintain provider keys and channels.

Why it matters

For indie creators and small studios, DramaClaw offers a self-hostable, model-neutral drama factory in place of a dozen disconnected tools.

For engineers, three design choices stand out: capabilities exposed over MCP, a gateway contract that hides provider differences, and long multi-stage jobs run with almost no infrastructure. Watch two things next: when the in-development agent and previz features ship, and whether the SaaS clause of the license ever opens.