Skip to content

[ commentary ]

OpenMuse puts an agent in a browser and a terminal

OpenMuse is an alpha, self-hosted agent template that browses, runs commands and keeps a task record you can inspect.

Published · on github.com · 2 min read

OpenMuse is an alpha, self-hosted agent template that browses, runs commands and keeps a task record you can inspect.

CopilotKit has published OpenMuse, a personal agent application that runs its own server, task worker and browser worker. The project is on github.com, where it is described as an alpha meant for self-hosting and for building on, under the MIT licence.

What OpenMuse actually contains

The core of it is what the repository calls an agent computer: a persistent Chromium profile plus an optional Linux workspace. According to the project description, the agent can read public pages, run commands inside its own container, work with files, and move PDFs between the computer and the app. A person can open the browser or the terminal and continue the work by hand. The repository notes that graphical desktops and autonomous checkout are still future work.

The second part is the record around the work. The activity surface is described as holding durable task plans, progress, input requests, pause, resume, cancel and retry, approvals, and saved receipts, with SQL leases used to recover interrupted work. Commands in the container are limited to 30 seconds and leave saved output and exit receipts. The interface is built with CopilotKit React Native for iOS, Android and web, and the repository invites teams to clone the template and customise it.

Where this touches office work

Two of the four jobs BARGO works on are visible here. Approval Runner is the closer match: an agent that asks for input, waits for a sign-off and stores a receipt is the shape of a request that has to be routed and recorded rather than answered in a chat window. Document Desk is the second: the PDF path in the repository moves an attachment into a form, produces a filled copy and a reviewed reply, which is the same sequence as an invoice or a contract being read and entered into a business system.

The limits matter as much as the features. This is an alpha. Live Google accounts and the richer thread features need their own configuration, and the deployment described in the repository uses a single owner behind a shared access key rather than multi-tenant authentication. The Docker computer and Google mail or calendar are not started by the default deployment blueprint.

A team wanting to test the idea could start with one narrow, repetitive request that already ends in a written approval, run it against fictional data first, and check whether the saved receipts give an auditor enough to follow. Whether that holds up on real supplier documents is the question to answer before anything goes near a production system.

Source: github.com — BARGO’s commentary on the linked source.

Reported in the AI Search newsletter:

Which task costs your team the most hours every week? Tell us, and we will tell you whether AI can take it and what it would cost.

← all posts · RSS