Build with AI

Copy

Build a web app that runs on Egg, a browser with a local AI gateway on the
user's machine (macOS and Windows). The app is a portable "Egglet": a web bundle
that talks to the local Egg Gateway through the official SDK. Follow this
contract exactly. Do not invent endpoints, API keys, or a backend.

SDK, imported from "@egg/sdk":
- Entry: export default defineEgglet({ async activate(ctx) { ... } }).
- The runtime discovers the local Gateway, runs the user authorization
  handshake, and gives you a scoped `ctx`. You never handle tokens or URLs.
- AI follows the browser Prompt API shape:
    const s = await ctx.ai.text.chat.create({ initialPrompts: [{ role: "system", content: "..." }] });
    const reply = await s.prompt("...");          // or s.promptStreaming("...") for chunks
- Data is SQL over your own tables, named with {{double_braces}}:
    await ctx.database.exec("INSERT INTO {{notes}} (id, body, created_at) VALUES (?1, ?2, ?3)", [id, body, Date.now()]);
    const rows = await ctx.database.all("SELECT * FROM {{notes}} ORDER BY created_at DESC");
- Outbound HTTP goes through ctx.fetch, never the global fetch.
- Mount the UI with ctx.routes.mount(() => import("./Page")).

manifest.json declares identity and what you use; the host enforces it:
  {
    "id": "notes", "label": "Notes", "icon": "notebook", "route": "/notes",
    "runtime": "portable", "delivery": "iframe",
    "tables": [{ "name": "notes", "columns": [
      { "name": "id", "type": "TEXT", "primary_key": true },
      { "name": "body", "type": "TEXT" },
      { "name": "created_at", "type": "INTEGER" } ] }],
    "uses": { "llm": true, "gateway": { "writes": true } }
  }

When Egg is not installed the SDK throws EggNotAvailable. Use the bundled helper
so the app degrades gracefully instead of crashing:
  import { runWithInstallPrompt } from "@egg/sdk";
  import module from "./index";
  await runWithInstallPrompt(module, { appName: "Notes" });

Build this: a single-page Notes app. A text box adds a note, stored with
ctx.database. A "Summarize" button sends the notes to ctx.ai.text and shows the
summary. Handle the not-installed case with runWithInstallPrompt. No backend, no
API keys, no database other than ctx.database.

Copy

Add Egg as the AI provider to a script that runs on the same machine as Egg
(macOS or Windows). Egg's local Gateway speaks the OpenAI Chat Completions API,
so use a standard OpenAI client and only change the base URL and key.

Discovery: read the Gateway's port and bearer token from its info file.
  macOS:   ~/Library/Application Support/com.eggbrowser.desktop/daemon_info.json
  Windows: %APPDATA%\com.eggbrowser.desktop\daemon_info.json
  shape:   { "port": <number>, "token": "<bearer>" }

Use it like OpenAI:
  baseURL = http://127.0.0.1:<port>/v1
  apiKey  = <token>
  POST {baseURL}/chat/completions  { "model": "", "messages": [...], "stream"?: true }

The model field is ignored; routing comes from the user's Egg settings. The
token is regenerated every time Egg restarts, so read it fresh at startup and
never hardcode it. Streaming responses are Server-Sent Events ending in
[DONE], exactly like OpenAI.

Build this: <describe your tool here>. Use the official OpenAI SDK for your
language, pointed at the base URL and token above.

https://eggbrowser.com/developers/llms.txt