OpenMuse is CopilotKit's personal-agent application. It combines chat, task progress, files, a browser and an optional Linux workspace. Its code is MIT licensed, and the project describes the current release as an alpha for self-hosting and building on. Official repository.
It is a separate project from Meta's Muse. Similar names do not establish an official relationship or feature parity.
I find the visible workflow more useful to inspect than the promise of an agent doing everything. Where is the plan? What survives a restart? Which action waits for review? Those details decide whether a system can be trusted with a real task.
What is included, and what still needs configuration?
| Layer | What to check |
|---|---|
| Application | Chat, saved tasks, plans, results and review states |
| Browser | Separate worker, persistent session and takeover surface |
| Terminal | Optional isolated Linux container |
| Real model | Provider configuration and its own usage cost |
| Gmail and Calendar | Google OAuth and reviewed write actions |
| Conversation persistence | Required CopilotKit Intelligence project key |
The repo explicitly says CopilotKit Intelligence is a separate service outside its MIT license. A key is required in every mode. This means MIT-licensed source does not make the whole runtime offline or free. Dependency explanation.
An open issue requests an optional alternative for self-hosting. Read its current status rather than assuming the requested change has shipped. Issue 73.
Start with the sample application
The README lists Node 24 LTS, pnpm 11.19.0 and an Intelligence project key. Its sample mode does not need a model, Google account or Docker. The documented initial commands are:
git clone https://github.com/CopilotKit/openmuse.git
cd openmuse
pnpm install --frozen-lockfile
cp .env.example .env
Use the project's current login/project-selection instructions to obtain the server-only Intelligence key. Set it privately in .env, then follow the documented server and web start commands. The README currently points to the web app at localhost:8081. Current quick start.
Do not paste keys into a prompt or publish your .env. Before changing from sample data to real accounts, read the current configuration and security notes in the repo.
Three things to inspect with fictional data
Start with a task whose outcome you can check without exposing a real account:
- Prepare a sample permission-slip response. Inspect the document, requested values, review step and saved outcome.
- Watch a built-in test page for a change. Check whether unchanged results create repeated alerts.
- Import sample transactions. Inspect the categories and totals rather than treating a polished chart as correct.
These follow the types of sample flows the project documents. They are suggested inspection steps, not tests I ran in this research pass. Sample tasks.
What the fresh Jev demo proves
CopilotKit's recent demo shows sourced comparison cards and clickable choices with Jev. The repository labels the aquarium-trip scenario as fictional and identifies scripted parts of the agent. It demonstrates a particular interface and decision flow; it does not independently establish unrestricted autonomous browsing or purchasing. Demo post, recording description.
When you build on it, retain the page behind each recommendation. A convincing card without a checkable source is still only a convincing card.
A deployment worksheet
Repository commit and version:
Sample or live mode:
Model/provider and usage budget:
Intelligence service dependency:
Browser worker and persistent storage:
Optional Linux computer:
Connected accounts and permission scope:
Actions that need owner review:
Backup location and recovery check:
Observed failure and retry behavior:
Estimated monthly runtime cost:
Self-hosting moves operating work to you. Check persistence, authentication and worker uptime before handing it an ongoing responsibility. Start with read-only research or drafts, then expand after checking the actual results.
FAQ
Is OpenMuse the open-source release of Meta Muse?
The cited repo identifies it as a CopilotKit project. Do not assume it is a Meta release.
Does MIT licensing mean zero running cost?
No. Separate services, model usage and hosting still matter.
Is every demo completely autonomous?
Read each recording's reproduction notes. The Jev scenario explicitly includes fictional data and scripted components.
Sources
- OpenMuse repository and current setup.
- Self-hosting dependency issue.
- CopilotKit's Jev demonstration.
Checked 1 October 2026. This is a source-based setup and evaluation guide, not a completed deployment review.