What is OpenMuse?
OpenMuse is CopilotKit's open-source personal-agent app, an open take on Meta's Muse. It is an alpha template, not a product to install: the README says to clone it and customise it, nothing is published as a package, and the repository has no tagged releases. The changelog's one entry is 0.1.0-alpha, dated 15 September 2026. A single codebase built with CopilotKit React Native serves iOS, Android and web, and talks to its server over AG-UI, an open event-based protocol for connecting agents to user-facing apps.
The agent gets a persistent Chromium browser you can take control of, an optional Docker Linux terminal with its own /workspace, and durable background tasks with plans, approvals and receipts. It can fill PDF forms from email attachments, watch public pages for changes or price thresholds, and read and draft in Gmail and Google Calendar, with every send or calendar change waiting for your review. Each deployment has one owner behind a shared access key.
Like OpenBot, it will not start without a CopilotKit Intelligence project key, which stores chat threads. CopilotKit lists it with its siblings OpenBot, OpenDots and OpenDuet. The repository was created on 15 September 2026 and has about 6,000 GitHub stars.
How much does OpenMuse cost?
| Plan | Price | What you get |
|---|---|---|
| OpenMuse | Free | MIT; self-hosted locally or with the Render Blueprint |
| Intelligence Free | $0 | 1,000 interactions per organisation a month hosted; local Developer Edition is non-production only |
| Intelligence Pay as you go | $49/month | 10,000 interactions included, then graduated rates; hosted or self-hosted |
| Intelligence Annual | Custom | Committed discount, offline deployment, custom SSO |
MIT licence. Requires CopilotKit Intelligence: Free ($0) hosted tier with 1,000 interactions per organisation a month; Pay as you go $49/month with 10,000 interactions included; Annual custom. You supply the model key and the hosting.
Is OpenMuse worth it?
OpenMuse shows what an open Meta Muse could look like, and it is a starting point for developers rather than an assistant you install. The repository was created on 15 September 2026, has about 6,000 GitHub stars and 866 forks, and has no tagged releases; the latest commits on main landed on 26 October 2026. Its own verification notes are candid: live model quality, live Google accounts and Intelligence replay across devices have not yet been accepted, and Android has been bundle-checked but not installed on a device.
The sandboxing is more deliberate than in most personal agents. The Linux terminal is off by default, and when enabled it runs in a non-root Docker container with a read-only root filesystem, all capabilities dropped, networking set to none, 512 MB of RAM, one CPU, 128 processes and a 30-second limit per command, with no fallback to the host shell. OpenClaw, by contrast, runs main-session tools on the host unless you configure sandboxing. The browser is a separate Playwright worker that allows only public HTTP(S) on ports 80 and 443 and rejects private addresses after DNS resolution. The docs say plainly that Playwright disables Chromium's own sandbox and that this is policy enforced by the application, not a kernel firewall. The optional E2B desktop provider is much looser: internet on, passwordless sudo, and every listening port in the VM reachable from the internet without app authentication.
Two dependencies point back to CopilotKit. Every mode, including the fictional sample, requires a CopilotKit Intelligence project key; the API refuses to start without one, chat threads are stored there, and Intelligence sits outside the MIT licence. Issue #73, asking for it to be optional, is still open. The free hosted tier covers 1,000 interactions a month, and Pay as you go is $49 a month with 10,000 included. Telemetry goes further than in OpenBot: the CopilotKit SDK sends runtime events to CopilotKit's PostHog sink at full sampling by default, and the CLI-issued CPK_TELEMETRY_ID lets CopilotKit join that usage to your Intelligence account. Web search is also on by default and sends model-written queries, with context from your conversation, to Parallel's free search service.
Against its neighbours, OpenMuse sits in the middle. Meta Muse is a free hosted product with WhatsApp access, a dedicated cloud VM and a Sentinel agent that approves outbound traffic, and it can buy things for you; OpenMuse keeps its database and files on your server, but autonomous checkout is future work and only its own email and calendar writes go through a stored review. OpenClaw reaches 20-plus chat apps with no paid dependency at all. OpenBot, from the same team, is built for organisations, with many Bots, company sign-in and a policy gateway that writes an audit row before every action. OpenMuse is single-owner, ships native phone and web apps, and carries an OpenBot adapter that is written and tested but switched off.
Where it earns its place
A React Native developer at a two-person startup
Ship a branded personal-assistant app without building chat, task and browser plumbing from scratch
Clone the repo, replace the capybara branding and result cards, and point MODEL at the team's provider. The code is MIT; threads need an Intelligence project, free up to 1,000 interactions a month hosted, then $49 a month for 10,000.
A self-hosting engineer who wants Muse without Meta
Have an agent read the school email, fill the PDF permission slip and draft the reply for approval
Deploy the Render Blueprint, connect Gmail through your own Google OAuth client, and approve each send; Google tokens are encrypted at rest on your server. Expect about $50 a month for two Standard Render instances before model and Intelligence costs, and set DO_NOT_TRACK=1 if you do not want usage reported.
A security engineer at a firm piloting personal agents
Check whether the agent's terminal and browser can reach internal systems
Read SECURITY.md and the browser worker README: terminal networking is set to none, and the browser rejects private and reserved addresses after DNS lookup. Keep the default Docker computer rather than E2B, whose VM exposes every listening port publicly.
Before you commit
- Setup
- Clone, run `pnpm install --frozen-lockfile` and `cp .env.example .env`, get an Intelligence key with `npx copilotkit@latest login` and `npx copilotkit@latest project select`, then `pnpm dev` and `pnpm dev:web` and open localhost:8081. Needs Node 24 and pnpm 11.19.0; sample mode needs no model, Google account or Docker. The browser worker, Docker terminal and Google OAuth are separate steps, and a Render Blueprint deploys the API, web app and browser together.
- Integrations
- OpenAI, Anthropic and Google models through MODEL=provider/model-id (the Render default is openai/gpt-5), with base-URL overrides for gateways and a Chat Completions mode for providers such as DeepSeek; an external AG-UI agent can take over conversation; Gmail and Google Calendar through your own OAuth client; Parallel's keyless Search MCP; optional E2B Desktop; a disabled, contract-tested OpenBot adapter.
- Data and privacy
- The database (embedded PGlite or PostgreSQL), PDFs, browser profiles and signing key stay on your host, and Google tokens are encrypted at rest. Chat threads go to CopilotKit Intelligence, search queries and context to Parallel, prompts to your model provider. CopilotKit SDK telemetry is on at full sampling and linked to your Intelligence account; COPILOTKIT_TELEMETRY_DISABLED=true or DO_NOT_TRACK=1 turns it off.
- Cost at scale
- On Render the API and the private browser each need a Standard instance at $25 a month with 2 GB of RAM, because the 512 MB Starter runs out of memory before binding a port, so hosting is about $50 a month plus 1 GB disks at $0.25 per GB. Add model usage and Intelligence: free hosted up to 1,000 interactions a month, then $49 a month with 10,000 included and graduated rates after.
- Limits
- One owner per deployment; terminal commands are non-interactive, capped at 30 seconds and 128 KB of output; the browser allows three active sessions and 20 saved profiles and disables service workers and WebSockets; no autonomous booking or checkout; scanned PDFs and OCR are unsupported; 13 open issues.
What’s good
- Terminal boxed in by default: no network, read-only root, 30-second commands
- One React Native codebase for iOS, Android and web
- Durable tasks with approvals, receipts and recovery after interruption
- Candid docs on what has and has not been verified
What’s not
- Will not start without a CopilotKit Intelligence key
- Usage telemetry on by default and linked to your Intelligence account
- One owner per deployment behind a shared access key
- Alpha with no releases; live Google and model use not yet accepted
Best for
Developers who want to build their own Muse-style personal agent on a phone and web app they control, and who accept CopilotKit Intelligence as part of the stack.
OpenMuse alternatives
OpenMuse: common questions
- Is OpenMuse worth it?
- OpenMuse is a carefully documented blueprint for an open Meta Muse, but it is a developer's template, it cannot run without CopilotKit's thread service, and it reports your usage to CopilotKit until you switch that off. We rate it 3.6 out of 5.
- How much does OpenMuse cost?
- MIT licence. Requires CopilotKit Intelligence: Free ($0) hosted tier with 1,000 interactions per organisation a month; Pay as you go $49/month with 10,000 interactions included; Annual custom. You supply the model key and the hosting.
- What are the downsides of OpenMuse?
- Will not start without a CopilotKit Intelligence key; usage telemetry on by default and linked to your Intelligence account; one owner per deployment behind a shared access key; alpha with no releases; live Google and model use not yet accepted.
- Who is OpenMuse best for?
- Developers who want to build their own Muse-style personal agent on a phone and web app they control, and who accept CopilotKit Intelligence as part of the stack.
- How much does OpenMuse cost for a team or in production?
- On Render the API and the private browser each need a Standard instance at $25 a month with 2 GB of RAM, because the 512 MB Starter runs out of memory before binding a port, so hosting is about $50 a month plus 1 GB disks at $0.25 per GB. Add model usage and Intelligence: free hosted up to 1,000 interactions a month, then $49 a month with 10,000 included and graduated rates after.
- Is OpenMuse safe for company data?
- The database (embedded PGlite or PostgreSQL), PDFs, browser profiles and signing key stay on your host, and Google tokens are encrypted at rest. Chat threads go to CopilotKit Intelligence, search queries and context to Parallel, prompts to your model provider. CopilotKit SDK telemetry is on at full sampling and linked to your Intelligence account; COPILOTKIT_TELEMETRY_DISABLED=true or DO_NOT_TRACK=1 turns it off.
- What are the best OpenMuse alternatives?
- The closest alternatives we have reviewed are OpenBot, OpenClaw, Magentic-UI (MagenticLite) and Browser Use.
Answers checked Oct 31, 2026.