Turn a rough idea into finished visual work — posters, product shots, illustrated articles, UI mockups, and short videos — in one local chat-and-canvas workspace.
Open Creative Agent (OCA) is an open-source, local-first multi-agent creative studio. You describe the outcome in plain language — goal, audience, copy, source material, constraints — and an orchestrator plans the work across 20 specialist agents for research, design knowledge, image generation and editing, video, articles, posters, and web/UI pages. Results land on a movable tldraw canvas next to the chat, ready to inspect, refine in follow-up messages, and download.
One git clone, one script, one browser tab. No Docker, no account, no cloud backend.
A real session at ~14× speed: search → design → generate → render → refine by conversation.
 |
 |
 |
 |
| You want to... | Go to |
|---|---|
| Run it on your machine in ~5 minutes | Quick Start |
| See what it can produce | Showcase |
| Understand what makes it different | Why Open Creative Agent |
| Configure API keys, models, ports | Configuration |
| Read the code, run tests, contribute | Development |
- 2026-07 🎉 Initial open-source release: local chat + canvas workspace, an orchestrator coordinating 20 specialist agents, and end-to-end generation of posters, product visuals, illustrated articles, UI pages, and videos.
- Brief-first, not prompt-first. Write the actual brief — audience, copy, product photos, style constraints — instead of compressing everything into one image prompt. The orchestrator decomposes it and routes each part to the right specialist.
- A real creative pipeline. Search → design reasoning → asset generation → editing → layout → rendering. Marketing sets, illustrated articles, and detail pages come out assembled, not as loose fragments.
- Text that is actually correct. Layout-heavy formats (posters, long images, knowledge cards, UI pages) are composed as web pages and rendered to pixels, so titles, dates, and body copy stay sharp and typo-free — the classic failure mode of pure text-to-image.
- Edit by conversation. "Center the title, make it 1.5× larger, drop the QR code" — follow-up messages refine the previous result instead of regenerating from scratch.
- Chat + canvas, side by side. Generated media appears on an infinite tldraw canvas you can arrange and compare, while planning and progress stream in the chat rail.
- Local-first and hackable. One FastAPI process serving a React UI. Readable Python on google-adk, pluggable model providers, no hosted dependencies.
Everything below was produced by OCA from a single brief (plus follow-up edits where noted). Click any image for full size.
|
Poster, refined over 4 turns "Make a poster for the 27th China Hi-Tech Fair — search for the details" → "center the title, brighter color, 1.5× larger" → "recolor the title, move body text down" → "remove the QR code". |
Amazon listing image set Main, multi-angle, lifestyle, and detail shots plus English copy — generated from one product photo, keeping the product identical. |
Reasoning-heavy generation Three trending outfits, each in front of a famous building from a different continent, under that city's real-time weather, annotated with a math, physics, and chemistry formula. |
Illustration series Four collage-style illustrations blending scenic landscapes into animal silhouettes, in bright high-impact colors. |
Mobile app UI A banking-app home screen: balance, card switcher, recent transactions, transfer / top-up / invest actions — professional and trust-first. |
Knowledge card 1080×1920 vocabulary card as a ready-to-open web page: cartoon illustration on top, phonetics, meaning, and example sentence below. |
Long-form vertical graphic A scrolling infographic of China's dynasties, each with a short introduction. |
Illustrated marketing article A Xiaohongshu-style product article with matching generated illustrations, written and composed end to end. |
Text-accurate event poster Exhibition poster with venue, dates, and organizer rendered exactly as provided — headline prominent, zero garbled text. |
Prerequisites: Python 3.14+, Node.js + npm, macOS or Linux, and a Google AI Studio API key (free tier works).
git clone https://github.com/GML-FMGroup/open-creative-agent.git
cd open-creative-agent
cp .env.template .env # put your GOOGLE_API_KEY inside
./scripts/start_local.shThen open http://127.0.0.1:9502 and describe what you want to make.
The first start takes a few minutes: it creates .venv, installs Python and web dependencies, builds the React UI, and launches the server. Later starts are much faster — see startup flags.
Gemini alone covers planning, writing, page generation, and Nano-Banana image generation. Optional keys unlock more tools — Seedream/Seedance images and video (Volcano Ark), web search (Tavily), and the GPT-backed article writer. See the provider table.
- Left — canvas. Generated images, rendered pages, and videos arrive as real tldraw objects: drag, arrange, zoom, compare versions.
- Right — chat. The brief, the agent's plan, expandable thinking steps, and progress live here. Attach product photos or reference images directly.
- Iterate in place. Every follow-up message edits the work in context.
- Download anything. Finished artifacts are served straight from the local app.
A single FastAPI process serves the API and the built React UI. Each chat request goes to an orchestrator that plans the task, then an executor dispatches steps to 20 specialist agents — web search, design knowledge, ad copy, text-to-image and image editing (Nano Banana / Seedream), reasoning-driven generation, video (Veo / Seedance), image understanding and OCR, background removal, illustrated articles, posters, reference-based page recreation, UI generation, HTML page generation, and HTML-to-image rendering. Session state lives in local SQLite; generated files stay on your disk.
Details, project layout, and validation commands: docs/development.md.
- Configuration — API keys and what each unlocks, model selection, ports, startup flags.
- Development — manual startup, architecture, project layout, tests, contribution notes.
The codebase is intentionally small and readable — one server, one UI, one agent package. Issues and PRs are welcome, and if OCA makes something you like, a ⭐ helps more people find it.
Directions we would love help with:
- More model providers — additional image, video, and LLM backends
- Windows support — a native startup path alongside the shell script
- UI polish and i18n — the workspace is young and moving fast