# Amuse Self-hosted personal AI agent: it researches the web, uses a live browser you can watch and take over, runs reminders and background checks, tracks goals and builds small apps you keep using, and asks before it sends, submits or pays for anything. Amuse runs on your own computer for one person. It needs an OpenAI API key and PostgreSQL. No account, hosted service or other API key is required. ## Features - Chat with an agent you name and shape, with side chats for separate topics - Web research with sources, using OpenAI web search - A live browser in the chat that fills forms and navigates sites, which you can watch, take over or stop - Reminders, monitors and a background heartbeat that only message you when something matters - Goals turned into step-by-step plans that the agent keeps working on - Documents, PDFs, images, podcasts and spaces: small apps with saved data - Memory you can read and edit, stored as plain workspace files - Approval cards for sends, form submissions, payments and new schedules, with once, task, site or always scopes - An encrypted vault for logins and API keys; the model only sees stand-in tokens - Voice mode with transcription and spoken replies | Hatch your agent | Ask for an app, get a working one | | ------------------------------------ | -------------------------------------------- | | ![Hatch](docs/media/hatch.gif) | ![Build an app](docs/media/build-an-app.gif) | | **It browses the web for you, live** | **Talk to it, and it talks back** | | ![Browser](docs/media/browser.gif) | ![Voice mode](docs/media/voice.gif) | ## Quick start with Docker Compose Requirements: Docker Engine with the Compose plugin, Git, and an OpenAI API key. ```sh git clone https://github.com/replit/amuse cd amuse cp .env.example .env # then set OPENAI_API_KEY docker compose up -d --build ``` Open **http://localhost:3000** and hatch your agent. Compose starts PostgreSQL alongside Amuse and publishes the app on `127.0.0.1` only. ## From source Use Node.js **22.9+**, pnpm **10.26.1** and PostgreSQL **14+**. ```sh corepack enable corepack prepare pnpm@10.26.1 --activate pnpm install --frozen-lockfile pnpm setup:browser # downloads Chromium for the agent's live browser cp .env.example .env # set OPENAI_API_KEY and DATABASE_URL pnpm build pnpm start ``` Open **http://localhost:3000**. The server migrates its schema automatically before listening. ## Your OpenAI key The agent (`gpt-6.2-sol`, with `gpt-6-luna` for quick steps), web search, image generation (`gpt-image-2`), transcription and spoken replies all run on your OpenAI key. Set a spend limit on your OpenAI project. Amuse also stops for the day at `AMUSE_DAILY_LIMIT` usage units (2,000 by default) and pauses background jobs until the next day. Models and every other setting are in [`.env.example`](.env.example). Besides OpenAI, Amuse only contacts the sites the agent visits for you and, for weather questions, the free [Open-Meteo](https://open-meteo.com/) API (weather data by Open-Meteo.com, CC BY 4.0). Fonts and every other asset are served locally. ## Scope Amuse is deliberately small. It's a web app for one person on one computer, it uses OpenAI models only, and it connects to services through API keys and its live browser rather than MCP servers. Notifications appear while an Amuse tab is open; there's no push to your phone yet. ## Security model Amuse is built for one person on their own computer, so it has **no login**. It listens on `127.0.0.1` only and refuses changes from other sites. Don't expose it to a network or the internet. The agent can only reach public internet addresses, saved credentials are only sent to the site you confirmed, and consequential actions ask you first. See [SECURITY.md](SECURITY.md) for the full model and [docs/AUDIT.md](docs/AUDIT.md) for licenses and checks. With Docker Compose, the browser runs without Chromium's own sandbox, and the container is the isolation boundary. Expect approval cards: the agent asks before it opens sites you did not name, before payments, and before it changes memory after it reads web content. The encryption key for saved credentials is created on first run in `.data/secret.key` (or the `amuse-data` Docker volume). Back it up together with your database; without it, saved credentials can't be read. ## Development ```sh pnpm dev # API on :3000 and the web app with hot reload on http://localhost:5173 pnpm typecheck pnpm test # server unit tests; no database, network or API key needed pnpm format # Prettier ``` The end-to-end suite (`pnpm test:e2e`) drives a real browser through every flow against a fresh database. It makes real model calls, so it spends API credits; see [CONTRIBUTING.md](CONTRIBUTING.md). The code is a pnpm monorepo: `apps/server` (Express API and the agent), `apps/web` (React and Vite), `lib/db` (Drizzle schema and migrations) and `tests/e2e` (Playwright). CI (`.github/workflows/ci.yml`) runs the format check, typechecks, unit tests, a dependency audit, the build, a Docker build and a secrets scan on every pull request. See [Contributing](CONTRIBUTING.md), [Security](SECURITY.md) and the [Changelog](CHANGELOG.md). ## Acknowledgements The workspace-file layout (`SOUL.md`, `IDENTITY.md`, `USER.md`, `AGENTS.md`, `MEMORY.md`, `HEARTBEAT.md`, `BOOTSTRAP.md` and daily `memory/YYYY-MM-DD.md` notes) follows [OpenClaw](https://github.com/openclaw/openclaw) (MIT). ## License [Apache License 2.0](LICENSE). Third-party dependencies retain their own licenses. Amuse is an independent project, not affiliated with or endorsed by OpenAI.