AGenUI Studio

中文 · Documentation · Changelog

AGenUI Studio is a self-hosted workspace for building, running, and iterating on agent-generated AGenUI interfaces. It includes a visual management console, a local SQLite knowledge base, configurable APIs/operators/rules, and native HTTP/SSE integration for production applications.

One real generation path: Contract-first intent, rule-driven design, data/operator binding, protected multi-turn editing, and a portable Runtime Package. The center is a public-safe Studio run backed by the included demo.offers source.

Real generated examples

Media, offers, pricing and a primary action	Inventory tags, original price and repeated rows	Hotel results with executable data binding
Media, separators, status, and coupon actions	Route, weather, alerts, and live status	Card detail without device chrome

The first four images are real Studio runs using configured models, rules, and Workspace tools, rendered by the included Renderer. The final image removes the device chrome so layout detail is easier to inspect. Together they exercise media, repeated lists, prices, labels, actions, and bindings. Original production references, data, assets, and internal rules are not distributed in this repository.

From intent to executable UI

The model decides meaning and design. Versioned Workspace tools constrain how the DSL is created, bound, and edited; the result is an expanded package that a Runtime can execute without reconnecting to Studio.

Reproduce a card from a fresh install

The gallery is not backed by a hidden template or custom application code. Start with the bundled data, then copy the public generation prompt, editing prompt, and small Markdown rule example. The first prompt creates a complete preview even before a source exists; the binder reports the mapping work required for a downloadable package. The editing prompt preserves existing data and actions while refining only the declared visual target set.

Use a permitted reference image only as visual input. Keep product-specific assets, business language, and private rules outside the repository; publish a small, generic rule revision when a visual system needs stable constraints.

What you can do

• Configure your own OpenAI-compatible or Anthropic Messages API endpoint and API key in the UI.
• Generate a card, preview it live, and refine it over multiple turns.
• Download the completed card as a self-contained Runtime Package whose data sources, bindings, actions and operators no longer depend on Studio IDs.
• Register a delivery target and publish that package from the workbench; callback delivery is built in and MQ adapters share the same delivery port.
• Manage knowledge, API definitions, operators, and design rules from the console; start with local demo data and replace any service when needed.
• Keep non-target components, bindings and actions protected during multi-turn edits.
• Embed the self-contained Card Runtime to execute packaged data sources, bindings, and TypeScript operators without a Studio connection.
• Publish parsed Markdown rules as immutable revisions with an atomic active pointer.

Model compatibility

AGenUI Studio supports user-provided models through OpenAI-compatible Chat Completions or the Anthropic Messages API. The following models have been tested by project contributors:

Model	Protocol	Status
Qwen (qwen3.8-max)	OpenAI-compatible	Verified
Claude Opus 4.7	Anthropic Messages API	Verified
GLM-5.2	OpenAI-compatible	Verified (text only)
Kimi K3	OpenAI-compatible	Verified
GPT-5.6 Sol / Terra	OpenAI-compatible	Verified

Results may vary by model version and provider. The tested GLM-5.2 configuration does not support multimodal input, and exact model IDs depend on the provider.

Architecture

Harness owns durable execution. AGenUI Agents use rules, catalog facts, knowledge, APIs, and published operators through bounded tools. The Workspace commits immutable artifacts for live rendering, download, or delivery through a callback/MQ adapter. Dashed blue boxes are replaceable ports, not required vendors or fixed implementations.

Local by default, replaceable by design

SQLite, BM25, Markdown rules, the Web renderer, and the bundled Runtime make the first run self-contained. Production deployments can keep the same contracts while replacing storage with MySQL, hot state with Redis, knowledge or data with MCP services, operators with a remote runtime, and delivery with a callback service or MQ consumer. The Runtime Package is portable: the bundled Runtime is a reference implementation, not a required hosting model.

Run locally

Requirements: Go 1.26.2+ and Node.js 20+.

cd web
npm install
cd ..
make dev

Open http://127.0.0.1:3100/admin/home. In Settings, enter your model endpoint, model name, and API key. Local data is stored under agenui-agent/var/ and is ignored by Git. make dev idempotently initializes the bundled knowledge/API/operator fixtures and starts the data service, Agent, and Web UI.

After a successful generation, use Download Runtime Package on the session page. Verify it independently with:

go run ./runtime/cmd/agenui-runtime /path/to/agenui-session.json

Integrate

The local demo and Agent intentionally share var/agenui/studio.db, so a TypeScript operator saved and published in the console is immediately available to the Operator Detail API and runtime adapter.

The console uses AGenUI management APIs. Applications can subscribe to native run events through:

GET /api/v1/sessions/{session_id}/runs/{run_id}/events
Accept: text/event-stream
X-AGenUI-Tenant-ID: <tenant>
X-AGenUI-User-ID: <user>

These identity headers are for the local/test profile only. Production must install an authenticated PrincipalResolver.

See Quick start, Card reproduction, Configuration, Architecture, Rule documents, Capabilities and acceptance, API reference, and Knowledge MCP.

Community & Contact

• GitHub Issues: bug reports and feature requests
• Email: [email protected]
• DingTalk Group: Chinese community
• WeChat Group: Chinese community

---

License

Apache License 2.0. See LICENSE. Third-party attributions are listed in THIRD_PARTY_NOTICES.md.

Security reports and contributions are welcome. See SECURITY.md and CONTRIBUTING.md.