English | 中文

AI Agent to generate web pages and safe APIs, AI generate UI once, API repeat everytime safely, quickly and stably.

Agent-to-API protocol and MVP demo: generate a simple task UI, tune an HTTP API request until it works,
then let users change filters, sort, and paging from the UI — which directly calls HTTP API
without going through the LLM again, no more token cost.

No SQL execution path. Writes or sensitive reads wait in the Admin approval queue.

• Node.js 18+
• Local APIJSONBoot-MultiDataSource (or compatible) at http://localhost:8080

cd ~/a2api
cp .env.example .env
npm install
npm test
npm run build
npm run dev

• Client (Vite): http://localhost:5173

• Admin (config approvals): npm run dev:admin → http://localhost:5174

  • Tables Apply + Call: run admin/sql/sys_Apply.sql and sys_Call.sql, reload Access/Request. If Call logs say GET denied for LOGIN, run admin/sql/patch_Call_access.sql (Access id 9003; 9002 is Apply).
  • Ordinary CRUD hits APIJSON HTTP directly; admin server only runs approve → Access/Request/Document/Chain
  • Admin tabs: Apply · Call logs · Stats

• API (Hono): http://localhost:3000

Open the client URL. Use Login (top-right) to open the account menu and set AI Model / Base URL / API Key (APIAuto-style). Try chips such as List the latest 3 moments with authors, then change sort/page and click Query / Refresh — the right panel shows usedLlm: false and the exact APIJSON body.

Curated chat examples live in conversations/; project Agent skills in .cursor/skills/.

Optional: set OPENAI_API_KEY in .env to refine bootstrap with an LLM. Without it, built-in intent rules for User / Moment / Comment still work (English and Chinese phrases).

Envelopes: { "version": "0.1", "<type>": { ... } }

• proposeRequest — candidate APIJSON call
• reviseRequest / decision — edit / approve|reject
• bindRequest — after code == 200, template + paramMap for UI-driven calls
• requestResult / status — outcomes

Read methods auto-execute. Non-sensitive post / put auto-execute with an audit row. Sensitive methods (default delete, override SENSITIVE_METHODS) wait for Admin Approve/Reject.

1. Bootstrap (chat / AI or rules) — generate UI + propose APIJSON → validate → execute until success → emit bindRequest
2. Steady-state (no LLM) — filter/sort/page → BoundExecutor merges paramMap into bodyTemplate → POST {baseUrl}/{method}

Top tabs:

• UI — chat bootstrap + bound table/detail/charts
• Data — APIAuto-style HTTP debugger
• Admin — sensitive approval queue + audit trail (auto_approved / approved / rejected)

Also:

• Embed APIAuto — iframe http://localhost:8080/api/index.html?send=true&type=JSON&url=...&json=... (share-link auto fill + send)
• Open APIAuto in new window — same share URL in a new tab

Agent / console automation:

a2apiAgent.switchTab("data")
a2apiAgent.debug({
  url: "http://localhost:8080/get",
  json: { User: { id: 38710 } },
  send: true,          // builtin send
  // useApiAuto: true, // or load iframe + auto send
})

export APIJSON_BASE_URL=http://localhost:8080
# or edit .env

Ensure the Demo schema is available on that server. Business layout tables are in opendoubao/sql/layout_demo_tables.sql (User / Moment / Comment plus Employee, Activity, Message, News, Notice, Blog, Article, Video, Music, Product, ShopOrder, Address, Category, …). After import, reload Access/Request. To add only categories, run opendoubao/sql/layout_demo_categories.sql; to add only addresses, run opendoubao/sql/layout_demo_address.sql (those pages also auto-import if the table is missing).

Writes (POST/PUT/DELETE): the Demo often requires a logged-in session (@role OWNER/LOGIN). The MVP still generates the request and shows the HITL Approve/Reject UI; if APIJSON returns "not logged in", log in via your Demo/APIAuto session cookies or relax Access for local testing. Reads work out of the box against the public Demo data.

npm test          # protocol + runtime unit tests
npm run build     # compile a2api + demo
npm run dev       # API :3000 + Vite :5173
npm run typecheck

Cross-device sync via database tables or file import/export — see the design plan.

We are always looking for more developers to help implementing new features, fix bugs, etc.
Fork the project and send a pull request.

https://github.com/TommyLemon

If you have any questions or suggestions, you can create an issue or send me an e-mail.

https://github.com/open-doubao-ai/OpenDoubao