PARAS MADAN
All writing

Latest in AI

Replit Amuse: What Stays on Your Computer and What Goes to the API

Check Amuse’s setup, model charges and approval boundaries before trying Replit’s self-hosted assistant with real work.

Replit released Amuse, an open-source personal AI assistant that runs on your computer. You run the application and its database, but the model still runs on OpenAI's API. If your reason for trying Amuse is to keep prompts off the internet, this isn't an offline assistant.

If your goal is to own the app and control where its workspace is stored, it's a more interesting option. I would start with a public research task and check three things before trusting it with real work: what leaves the computer, which actions ask for approval, and what happens when the app isn't running.

Check the two deployment requirements

The repository offers a Docker Compose setup and a source install. Compose brings up PostgreSQL with the app and exposes the interface on your computer at localhost:3000. The source route requires Node.js, pnpm and a PostgreSQL instance. Both need an OpenAI API key, and neither requires a Replit account.

Read the repository and deployment files before running them. Keep the service on the same machine and use a throwaway workspace for the first trial. The project is designed for one person and has no app login; it isn't a team workspace to put on a network. The security notes also say Compose uses the container as its browser isolation boundary rather than Chromium's own sandbox. That matters if you plan to visit sites beyond your first known source.

Trace one task from the browser to the model

Use a public question with a known answer, such as: “Read this project's public release notes and tell me which version added CSV export. Include the source. Don't sign in, send anything, or save a memory.”

The local application keeps its workspace data under your control. The model calls and built-in web search still go to OpenAI. If you use voice, transcription and spoken replies also rely on the API. If the task uses the live browser, the sites it visits receive normal browser requests. These are different destinations, so don't treat “self-hosted” as a single answer to where information goes.

When you test a later workflow involving a document, decide whether the relevant excerpt may be sent to the model before providing it. No change to where the database sits will make that inference happen offline.

Check approvals with a harmless case

Amuse has approval prompts for consequential actions and makes the live browser visible. Use a public task with no account session or credentials and inspect any approval prompt before continuing. Check that the site and action match what you meant. Prefer a one-time approval while learning how the app behaves.

If a proposed action goes beyond the task, decline it and see whether the app can still answer from the source you approved. That tells you more than a demo that presses through every confirmation. I wouldn't add a saved login just to check whether public research works.

Budget the API separately

The app is open source, but the work done by the model isn't free. Replit's repository lists GPT-6.2 Sol as the main model and GPT-6 Luna for faster steps. Use OpenAI's API pricing to check the rates for the features you actually try, then inspect the resulting usage in your project.

Amuse also has a daily limit in application usage units. Those units aren't a dollar estimate. Run the same representative research task a few times before deciding whether reminders, web search or voice need their own budget. Leave background jobs off while you're isolating the cost of a single task.

Choose a job the local app can reliably finish

For the first week, use Amuse for something you'll review when you return to the same computer. The repository says notifications appear in an open Amuse tab and don't push to your phone. A reminder that only matters if you see it at a specific time deserves a delivery test before you depend on it.

I would use Amuse first for reviewing public information, not for reminders I can't miss or work that requires private credentials. The choice is about which parts you want to run yourself. If your hard requirement is local inference, changing the app host doesn't meet it.

I'm Paras. I write about AI products and how teams use them.

If you're releasing a developer tool and want to explain its practical tradeoffs clearly, I can help with the writing.

Book a 30-minute call.