Meta Muse: A Practical Guide

[ikaijua]

Introduction to Meta Muse (long-lived wiki page — contributions welcome; compiled 2026-09-22).

1. What is Muse?

The three-sentence version:

Muse is Meta's personal AI agent — you give it a task ("book this trip", "reply to that email") and it opens a browser, fills forms, compares prices, places orders, and updates your calendar, then reports back. It is not built for chatting; chat is just one way to assign tasks.

What is an agent? A regular AI answers questions; an agent does work — it breaks a goal into steps, operates software, handles hiccups, and you only review the result. Muse is the first of its kind to top the US App Store free chart (about 10 days after launch, overtaking ChatGPT).

• Launched September 8, 2026 (internal codename Hatch), US users only for now (18+)
• Runs on Muse Spark 1.3, Meta's proprietary model — note: not the open-weight Llama family
• Entry points: iOS / Android apps (muse.ai), the web, and WhatsApp (command it like texting a friend)
• Market reaction: 编程工具可以添加aws提供的codewhisper #1 on the US App Store free chart ~10 days in; Meta stock jumped 11.43% in one day (+$190B market cap)

2. Quick start

1. Install the Muse app from the US App Store, or open muse.ai (requires a US network environment; unavailable in mainland China)
2. Log in with a Meta account
3. Connect "connectors" as guided: email, calendar, payments — this step decides what Muse can do; it can only use what you authorize
4. Give the first task in plain language, e.g. "find me flights to New York next Wednesday, arriving in the afternoon, cheapest first"
5. Review: Muse shows what it did; sensitive actions like payments and sending messages trigger an approval card that only you can confirm

The biggest gotcha is permissions and trust, not features. To work for you, Muse must sign into your services — and its requested data scope is on the larger side among consumer agents. Days before launch, Meta agreed to an $18B multistate settlement over social-media harms, making privacy its weakest flank. Recommendation: start with low-risk connectors (e.g. calendar), get comfortable, then gradually enable email and payments.

3. What it can do

• Multi-step errands end-to-end: travel booking, form filling, shopping, schedule management — across many sites and apps
• Email on your behalf: drafts, replies, sends in your style (approval card before sending)
• Spends money safely: pays via Stripe Link one-time virtual cards; your real card number never reaches merchants
• Keeps working in the background: close the app and the task continues; you get notified when done
• Operates a browser: clicks, scrolls, and fills web forms like a human, inside an isolated VM
• Command it from WhatsApp: no separate app needed
• Creates documents: pulls material from Dropbox / Box / SharePoint (mid-September update)

4. Typical use cases

• Travel end-to-end: "Arrange my Austin trip next month: fly out Tuesday, back Thursday, book flights + hotel, add to calendar, and email the itinerary to the team"
• Shopping proxy: "Buy me an air fryer under $100 with 4+ star reviews" — it compares options and checks out (you approve the one-time card)
• Admin chores: post a second-hand sale, book a DMV appointment, cancel a subscription, fill an insurance claim
• Inbox triage: daily pass over your mailbox with drafted replies you confirm in one tap
• Personal errands: "Remind me to order a birthday gift for my aunt before Friday — gardening stuff, under $50"

5. Advanced: the security architecture (why you can let it spend)

Muse is the most transparent consumer agent on safety design. Core idea: surrogate tokens + host-side approval.

• Isolated VM: the agent runs in a dedicated systemd-nspawn virtual machine; it cannot touch your real system
• Surrogate tokens: when connecting services like email, the agent gets substitute tokens and never sees your real passwords or OAuth credentials
• Sentinel: an independent approver on the host machine — the sole authority for connector actions and network access
• eBPF data-flow tracking: sensitive-data flows are monitored at the kernel level
• One-time payment cards: checkout via Stripe Link single-use cards; merchants never see your real card
• Big bug bounty: up to $300K, including $130K specifically for prompt injection (web text tricking the agent into overreach) — which Meta openly calls an unsolved problem
• New "direct network" option: TCP/UDP access can be opened, but every connection raises a hostname-and-port approval card

In short: it can act for you, but every payment, message, or unusual access stops at an approval card — unless you blindly approve everything.

6. Pricing

Tier	Price	Notes
Free	$0	~100M tokens/week (official figure); enough for light daily use
Power	$20/mo	More usage
Maximum	$100/mo	Heavy use
Always-on	hosting plan	Keeps the agent running 24/7

An invite-code system grants extra free quota during the early rollout.

7. FAQ

Q: How is it different from ChatGPT / Claude?
ChatGPT and Claude are "conversation + tools" — you steer step by step. Muse is "delegate + review" — you state the goal once, it works, you check the result. Complementary, not a replacement: hard decisions still belong to ChatGPT/Claude; errands go to Muse.

Q: Can I use it from China?
No. US-only (18+) for now — signup, network, and payment all have US barriers. See the comparison in section 8 for alternatives.

Q: Can it spend or send things without asking?
Not by design: one-time cards for payments, approval cards for messages/orders. The real risk is you — bulk-approving without reading hands over the keys.

Q: How much is free?
~100M tokens/week (official figure); invite codes stack extra quota.

Q: Is there an Android app?
Yes — iOS / Android / web / WhatsApp. But note Android performance is far weaker than iOS (Google Play ranking once ~338, while iOS hit #1).

Q: Should Meta's privacy record worry me?
It is the biggest controversy: Muse wants broad data authorization, and Meta just paid an $18B settlement. The technical architecture (VM + surrogate tokens + approval cards) is thoughtfully done; company-level data trust only time will tell. Keep human confirmation on for sensitive tasks.

8. Alternatives at a glance

	Interface	Authorization	Pricing	Best for
Muse	App + WhatsApp	Surrogate tokens + approval cards (most conservative)	Free–$100/mo	US users who want the safest design
Instinct	SMS / phone	Full credential delegation (max power)	Not public, invite-only	Users willing to delegate everything
Town	Email-native	Passwordless, scoped OAuth	$15–$199/mo	Work scenarios
Gemini Spark / Claude Cowork	Their ecosystems	In-ecosystem connectors	With subscription	Users already in Google/Anthropic ecosystems

---

Official references first:

• Muse
• DataCamp: Meta Muse — The Personal AI Agent Built for Everyone
• daily.dev: what Zuckerberg said about privacy, pricing
• Carly: 7 Best Meta Muse AI Alternatives in 2026
• fabric.so: Instinct vs Meta Muse