Home input & Skill pick
Attachments, model, and skill configured in one go; the running queue shows local engine slot status.
Local-first · Infinite canvas
One sentence in, storyboard to film out.
BigBotArt puts video and image generation agents inside an infinite canvas: a persistent chat on the right edits nodes directly, while the storyboard, shot takes, and timeline export close the loop in one project. The h3.c video engine renders with sound, the mlxgen image engine downloads on demand, and every computation stays on your Mac.
Image gen · Storyboard still
Z-Image · Paper collage
Shot 02
Agent re-run · h3.c queue
Timeline entry
Selected Takes ×3 · sound straight out
Agent · Chat
Build a scene on the director's board: steppe, a wide river, the heroine leaps across… paper-collage style, under 60s.
Plan placed on canvas · Paper-collage drama Skill injected.
ToolRun 7 tools · 6 done · 1 running
storyboard.materialize → run_take · shot 2 h3.c queue
Interactive demo
Three main interfaces, one creative pipeline.
Sample project “River on the Steppe · Short play”: start on Home → arrange on the canvas → export from the timeline. Click the tabs to switch, or pause the auto-play.
Local-first · All data stays on your Mac
What would you like to create today?
Build a scene on the director's board: the two leads walk across the steppe and meet a wide river; the heroine leaps across while the hero paces the bank. Paper-collage style, under 60 seconds.
References / files
local LLM
Paper-collage drama
Planning agent
Generate
Asset · Heroine v3Character
Doc · Brief
River on the Steppe
Scene: steppe / wide river · paper collage
Image gen · Storyboard still
mlxgen · Z-Image-Turbo · 832×480 · seed ×4
Audio ref
Shot 01 · Heroine leaps
h3.c · FL2VA · 1280×720 · sound
Take 1Take 2Take 3 ✓
Agent re-running · 06:12 left
Shot 02 · Hero paces
Shot 03 · Planned
To generate · queued
Timeline / edit
68%
00:04 / 00:52
Take pool
Shot 01 · Take 3
Sound out · continuity ✓
0s
8s
18s
32s
48s
Video
Shot 01 · 8s
Shot 02 · 10s
Shot 03 · 8s
Shot 04 · 12s
Audio
h3.c synthesized track · sound straight out
Capabilities
Canvas, agents, and assets — designed for video production.
Infinite canvas workbench
11 node types cover storyboards, shots, assets, and workflow I/O. Edges carry semantic labels, the inspector renders a form per type, and any change marks nodes stale while syncing the parameter bar.
Persistent agent chat
Tool calls collapse into foldable ToolRunCards, and rows with a nodeId jump straight to the canvas node. Markdown storyboards and ask.user confirm cards align with the worker protocol.
Two-tier asset library
Project assets and the global library share the work: categories, tags, reference tracking, and a recycle bin. Drag a card onto the canvas to spawn an asset node; outputs archive back to the library.
Typical scenario
E-commerce product films: from Skills to timeline.
Describe the shots in one sentence on Home and pick the “E-commerce” Skill to inject its brief. The agent lays storyboard and shot nodes on the canvas, batch-generates takes, and you drag them into the timeline — FFmpeg exports the final film with sound.
- Skills library: canvas templates plus instruction sets, injectable into the current project
- Workflow mode: auto-completes entry/exit nodes, validates with validateWorkflow, and runs in topological order
- Timeline: take pool, dual audio lanes, continuity checks, and export progress
Closed loop
From brief to export, four steps on one screen.
Canvas arranging & agent run
Storyboard nodes, shot nodes, and asset references edited visually on one infinite canvas.
Take picking & timeline editing
Drag takes into the tracks; video and audio align on dual lanes, with an interactive playhead and trims.
Local FFmpeg export
Presets, progress, and continuity results stay on your Mac — nothing uploads to the cloud.
Local engines
Models and compute stay on your Mac.
The first-launch wizard probes the environment, installs components, and downloads models. Settings manages LLM sources, the provider registry, and queue mutex in one place, while VRAM estimates help you trade off local versus cloud.
h3.c video engine
MiniMax-H3 (BF16) shot-level video with audio: text-to-video, first-last frame, and Ref2VA; optional offline LoRA style packs. Built for product spins, scene extensions, and storyboard previsualization.
mlxgen image engine
Package mlx-gen-jzg: downloads Z-Image-Turbo, the Qwen-Image family, SeedVR2, and more on demand for reference images, first frames, and asset nodes. Linked to the canvas parameter bar — staleness after edits is always visible.
Start your first agent project on Mac.
The macOS Apple Silicon build is out (v0.1.1). Install it, then finish first-time setup with the engine onboarding wizard. Source code and changelogs live on GitHub.
Get for macOSFAQ
What you might want to know about BigBotArt.
The six questions below are rendered as plain text for AI and for visitors: every answer is a complete, self-contained statement you can quote on its own.
What is BigBotArt?
BigBotArt is a local-first macOS workbench for AI video and image generation agents: the agents live inside an infinite canvas, a persistent chat edits nodes directly, and storyboards, shot takes, and timeline export close the loop in one project.
Which systems are supported? Do I need to be online?
BigBotArt ships as a macOS Apple Silicon (arm64) app (v0.1.1). Models and compute run on your own Mac and assets never need to be uploaded to a cloud; you only need a connection to download models the first time and to update the app.
Which models and engines are used?
The h3.c video engine runs MiniMax-H3 (BF16, ~144GB) for shot-level video with audio — text-to-video, first-last frame, and Ref2VA; optional style packs merge LoRA offline into a copy of the base model. mlxgen (package mlx-gen-jzg) downloads image models on demand: Z-Image-Turbo (default T2I/img2img), Qwen-Image-2.1 and Turbo (with references), Qwen-Image-Edit (instruction edit / inpaint / outpaint), FLUX.2 Klein 9B (multi-reference / outpaint), SeedVR2 3B (upscale), and a Qwen3-VL caption model for Agent vision; an Uncensored T2I variant needs a manual local convert. Agent chat uses any OpenAI-compatible endpoint (for example LM Studio). Final export runs through local FFmpeg; optional post-process Real-ESRGAN / GFPGAN / CodeFormer must be installed separately.
What does the end-to-end workflow look like?
Four steps on one screen: ① describe the brief and pick a Skill on the home screen; ② arrange storyboard and shot nodes on the canvas while the resident agent executes; ③ pick takes from the take pool and drop them into the timeline; ④ export the finished cut with local FFmpeg.
What are Skills?
A Skill is a preset in two forms — a canvas template plus operating instructions — that you inject into the current project in one click, so the agent lays out storyboards, locks the style, and runs batch generation the way you want.
How do I download it? Where is the source code?
Download the macOS Apple Silicon build from GitHub Releases: https://github.com/jiangzhenguo/bigbot.art/releases/download/v0.1.1/BigBotArt-0.1.1-arm64.dmg . Source code and changelog live in the same repository, https://github.com/jiangzhenguo/bigbot.art , and you can also reach us at jyllyj826@gmail.com.