11/5/26

Free Maestro v4 connects to ComfyUI and AI agents

Soft

Blaine Brown (@blizaine) announced in his post on X the release of the fourth version of Maestro. This is his local AI studio, where video, images, music and editing are brought together in a single app. According to the author, v4 adds a full-fledged MCP server, ComfyUI integration and support for the Pika H3 Flex model, and the tool itself is available for free.

For anyone assembling videos out of generative clips, the main thing in the update is the links to other tools. Judging by the list of new features, Maestro can now be connected to an already configured ComfyUI and hand part of the control over to an external AI agent via the MCP protocol. The post itself has no details — the author promises them in the continuation of the thread — but a description of both features is already in the project's repository.

“Maestro is the most intuitive, powerful, local, open creative AI studio available for free,” is how Blaine Brown described his project.

That is the author's own assessment. Along with the announcement he showed a video on the theme of VisionQuest and Avengers Doomsday, put together in Maestro v4.

What the studio can do

Judging by the Maestro repository on GitHub, the project installs in one click, and local models run on the user's computer. External AI services and remote access are connected optionally. It is built on the WanGP pipeline. The app is split into three parts:

Director builds a plan for a music video or a short film from a song or a storyline. It prepares character references and key frames, generates clips and assembles the result. In Auto mode production runs from start to finish; in Manual mode every stage can be checked. Plans, prompts and takes are kept in a single project panel, so individual shots can be reworked one at a time.
Studio handles video generation from text and references. Here video can also be animated, extended, blended and corrected. The studio creates and edits images, music, speech, cloned voices and sound effects. Among the models named are MiniMax H3, LTX, Flux, Qwen, MOSS and Fish Audio, and there is a task queue. The project is optimized for LTX-2.5 and MiniMax H3.
Editor is a non-destructive multitrack editing desk. It has trimming, transitions, speed, volume and canvas settings. The finished production from Director lands here, and a shot can be sent back for regeneration.

According to the description of version 4.0.0 dated November 4, 2026, the editing desk has been expanded: reviewable auto-edit plans with undo, chroma key and export to ProRes 4444 MOV with transparency have appeared. Animated titles can be described in words right in Studio or in Editor.

ComfyUI and AI agents

The ComfyUI integration, according to the repository, makes it possible to use already configured model folders, import and validate ComfyUI API graphs and route supported Studio and Director requests to an existing ComfyUI server. The feature is optional and disabled by default; nothing is enabled or installed automatically. ComfyUI itself and its custom nodes are installed and maintained separately.

The built-in MCP server lets compatible agents find out what the studio can do, write prompts that take the selected model into account, use saved characters and references, send tasks to Studio and Director and manage projects. An agent can also bring its own music video plan, while Maestro will analyze the song itself. You can connect locally on the same computer or from another machine via Tailscale with authorization; permissions are granted to the agent within set limits.

Six steps instead of fifty

Another notable new feature is support for Pika H3 Flex. In Maestro v4 it is marked as experimental: generation from frames and from references, handling of the source audio and five resolution levels, including an experimental 1080p. According to the ComfyUI documentation, Pika H3 Flex is a fine-tuned version of MiniMax H3. The generator has been distilled from 50 steps down to six. The model ships as two LoRA adapters on top of the base MiniMax H3, so no new architecture is needed. A five-second clip at 1344×768 is computed in roughly a minute on a single graphics card. Without distillation, H3 requires 20 to 50 steps.

Other properties of the model described in the ComfyUI documentation also matter for editing shots:

dialogue, effects and music are generated in stereo in the same pass as the picture;
up to 12 references can be fed into a single generation: up to 9 images, 3 videos, including a motion reference, and 3 audio clips with voice samples;
cuts, camera movement and lines are described in a single prompt, meaning the model can handle multi-shot staging;
the weights of both adapters are open.

The adapters are distributed under the MiniMax H3 Community License Agreement. The documentation advises studying its terms, including territorial restrictions and commercial use, before putting the result into commercial work.

What is still missing

Maestro's own license is not stated in the post, and the repository refers to the licenses of the individual models. A studio planning to use generated material in commercial projects will have to separately check the terms of Maestro, the WanGP pipeline and every model it connects. The author did not name specific minimum requirements for the graphics card and memory. The repository describes an Auto-Tune feature that picks starting settings based on the graphics card, the amount of video memory and RAM, while quantization and offloading to memory should help on weak GPUs.

For comparison, one can look at the separate MCP server from Comfy-Org. Through it, Claude Code, Claude Desktop, Cursor and other agents run graphs in a local ComfyUI, track the task queue, collect video previews into a contact sheet, check which nodes and models are installed and make variations of a single graph. It is a different product, but together with Maestro it probably reflects a common direction: routine work such as launching generations and sifting through variants can increasingly be handed over to an agent. In this logic Maestro goes further than the graph — the agent is given access to staging planning in Director and to working with projects.

Which of the new Maestro v4 features would be more useful to you?