Skip to main content
Back

Replit Open-Sources Amuse, a Personal AI Agent Built in Three Days

Replit's agent swarm cloned a popular personal AI assistant in three days, releasing Amuse as a free, self-hosted alternative that keeps your data local.

Repo, Full Access
  • Replit open-sourced Amuse, a self-hosted personal AI agent under Apache 2.0.
  • Built in three days by Replit Missions, its new agent swarm product.
  • Features live browser automation, reminders, voice mode, and small app building with approval cards.
  • Runs locally on 127.0.0.1 with PostgreSQL and your own OpenAI key, no account required.
  • Memory stored as plain editable markdown files following the OpenClaw convention.
  • Scope limited to one user on one machine, no mobile push notifications yet.

Replit has released Amuse, an Apache 2.0-licensed personal assistant that stores its application data, memory and credentials on the user's computer. The app still requires network access: model inference and web search use OpenAI, while its browser agent connects to sites on the user's behalf.

Replit says it built Amuse in three days by giving Replit Missions, its agent-swarm product, a single prompt to replicate a leading personal AI assistant. The company then used its Security Scanner to identify and patch vulnerabilities before publishing the code. The result gives developers both a self-hosted assistant to inspect and a public example of what Missions can produce.

What Amuse can control

Amuse combines chat, browser automation, scheduled work and editable memory in a single-user web application. Actions that send information, submit forms or initiate payments require approval.

Capability How it works
Conversations Users name and shape the main agent, then create side chats for separate topics.
Web research OpenAI web search gathers information and returns sources.
Browser automation A live browser navigates sites and fills forms inside the chat. The user can watch, take control or stop it.
Background work Reminders, monitors and a heartbeat loop run scheduled checks and surface results.
Generated content The agent can create documents, PDFs, images, podcasts and “spaces,” which are small applications with persisted data.
Memory Preferences and history remain available as editable workspace files.
Secrets An encrypted vault stores logins and API keys while exposing placeholder tokens to the model.
Voice Voice mode supports transcription and spoken responses.
Amuse onboarding screen for creating a personal agent
Amuse guides the user through naming and configuring a personal agent.

The laptop becomes the trust boundary

Amuse assumes one trusted user on one trusted computer, so it has no login screen. The server binds to 127.0.0.1, rejects cross-origin state changes and limits the browser agent to public internet addresses. Any process or person with access to the host remains inside that security boundary.

Saved credentials are restricted to the domain the user confirms. The model receives stand-in tokens instead of raw secrets, and consequential actions generate approval cards. Permissions can apply once, for the current task, for a specific site or permanently.

Amuse live browser displayed beside a scheduled tasks panel
The embedded browser exposes the agent's actions and allows manual takeover.

These controls reduce the exposure created by automated browsing, but websites can still contain prompt-injection attempts and users can approve harmful actions. Host security, domain restrictions and the approval flow therefore remain part of the operating model.

Memory stays readable

Amuse follows the OpenClaw workspace convention, storing its configuration and memory in files such as SOUL.md, IDENTITY.md, USER.md, AGENTS.md, MEMORY.md and HEARTBEAT.md, alongside daily notes. Users can inspect, edit, delete or copy those files without relying on a server-side memory interface.

The file layout also makes agent behavior easier to audit. Identity, instructions, long-term memory and scheduled work occupy separate documents rather than an undocumented application database.

From clone to localhost

The repository is a TypeScript pnpm monorepo containing an Express server, a React and Vite frontend, and Drizzle backed by PostgreSQL 14 or newer. Docker Compose supplies the shortest setup path:

git clone https://github.com/replit/amuse
cd amuse
cp .env.example .env
# Add OPENAI_API_KEY to .env
docker compose up -d --build

After the containers start, the interface is available at http://localhost:3000. The default model configuration and usage limit are:

Purpose Default
Primary agent gpt-6.2-sol
Quick steps gpt-6-luna
Image generation gpt-image-2
Daily limit AMUSE_DAILY_LIMIT=2000

The daily limit pauses background jobs after the application reaches its configured usage-unit threshold. It constrains unattended activity from reminders and the heartbeat loop, while OpenAI billing continues to depend on actual API usage.

Amuse doubles as a Missions demo

Replit Missions accepts a goal, budget and operating boundaries, then assigns agents to planning, implementation, testing and iteration in parallel. A separate AI judge evaluates the work against criteria set in advance, and the process stops when it exhausts its budget.

Amuse provides an inspectable output from that workflow rather than a closed demonstration. Developers can review the architecture, run the code, test the security assumptions and fork it under Apache 2.0. Replit's reported three-day build remains a project account, not a general benchmark for agent productivity.

Limits before connecting real accounts

  • The repository is new and small, so its reliability and maintenance record remain unproven.
  • Notifications require an open browser tab, and the current release has no mobile push support.
  • The application targets one person on one computer. Its lack of authentication makes it unsuitable for a shared or internet-facing deployment.
  • Under Docker Compose, Chromium runs without its own sandbox. The container provides the isolation boundary.
  • Port 3000 should remain bound to the local machine rather than exposed to the public internet.
  • Prompts and model inputs travel to OpenAI, while browser traffic reaches the sites the agent visits. The deployment is self-hosted but depends on external services.
  • Stored credentials and automated browsing increase the impact of prompt injection, mistaken approvals and compromised host access.

Amuse gives developers a compact, forkable implementation of a personal agent with visible browser activity, editable memory and explicit approvals. Its practical scope is narrow by design: one trusted user, one trusted machine and an OpenAI-backed agent operating behind a local interface.