Meta's Muse Opens to Developers: A Field Guide to Agentic Commerce

Meta Connect 2026 opened this week in Menlo Park with Mark Zuckerberg talking up 'personal superintelligence' for glasses and wearables. But the more concrete story for builders had already landed ten days earlier: Meta's consumer AI agent, Muse, got a developer platform — and it's a useful field guide to how a major platform is choosing to wire LLM agents to real payments.

What Muse actually does

Muse launched on September 8, 2026 as a personal agent available on the web, iOS, Android, and WhatsApp, with glasses integration flagged as next. It can send emails, book travel, negotiate bills, fill out forms, turn a recipe into a shopping list, and complete purchases at checkout using Link by Stripe. Pricing runs a metered free tier plus $20/month and $100/month plans. Meta says each session runs inside an isolated 'Secure VM,' watched by a separate monitoring agent it calls Sentinel, and that Muse 'won't have visibility into people's passwords or payment methods.'

The connector platform is the real news

On September 18, Zuckerberg opened Muse to third-party developers via muse.ai/platform. Any team with a public API can submit a connector — a description of what the service does and how someone would ask for it — which Meta reviews across three passes (functional, security, legal) plus end-to-end testing before it goes live.

You bring the API — Muse brings the agent, the browser, and the context of what the person actually wants. People reach your service just by asking for it, and their agent takes it from there.

— Mark Zuckerberg

What's worth studying, not copying

• The review gate applies an app-store model to agent actions, not just agent code — a pattern worth considering for any team shipping a tool-calling agent with payment access.
• The isolated-VM-plus-supervisor split is a legible guardrail architecture: the agent that acts is not the same process as the one watching it act.
• Meta's 'no visibility into passwords or payment methods' line is Meta's own description of its system, not an independently audited spec — a design goal to interrogate before anyone leans on a similar trust claim.

None of this requires Meta's stack to be relevant. Teams building trading bots, support agents, or fintech workflows are increasingly answering the same three questions Meta just answered in public: how do you scope a connector's blast radius, who signs off before it ships, and what can you actually prove the agent never sees.