Amazon Blocked Muse, Google Sold It a Key: The Agent Access Wars Have Two Playbooks

Background

Meta launched Muse in September 2026, a personal AI agent that shops on users' behalf, pays via single-use virtual cards issued through Link by Stripe, and manages email and calendars. Within days, Amazon blocked Muse from Amazon.com with a warning that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use," echoing its earlier standoff with Perplexity's shopping agent. Almost simultaneously, on September 16, Google launched Home MCP in early access, letting Claude, ChatGPT, Hermes, OpenClaw, and Google's own Antigravity read device states, control Nest cameras, doorbells, thermostats, and Matter devices, and analyze camera event history, all gated behind the $20/month Premium Advanced tier, US only, via OAuth.

Key Findings

The Fight Is Over Access, Not Capability

The prevailing framing of the agent wars is a model race: whose agent is smarter, faster, more autonomous. The Muse and Home MCP announcements, landing days apart, show that framing is wrong. Muse is capable. It can navigate Amazon, add to cart, pay through Stripe-issued single-use cards, and manage a user's inbox. None of that mattered once Amazon flipped a switch. Meanwhile Google made the opposite move on a different surface, letting five agents (including direct rivals to its own Antigravity) into the smart home graph. Capability is table stakes and commoditizing on a quarterly cadence. Access to the surfaces where agents actually act (shopping carts, payment rails, camera feeds, thermostats) is scarce, defensible, and now being priced or denied by its owners.

Amazon's Block Is a Pattern, Not an Incident

One block could be a dispute. Two is a policy. Amazon told GeekWire it asked Meta to keep Muse off Amazon.com and blocked the agent only after those requests failed, citing three specifics: Muse did not identify itself as an AI agent, it allegedly captured user credentials, and it scraped order history while operating inside users' sessions. Amazon anchors the block in Conditions of Use language prohibiting "data mining, robots, or similar data gathering and extraction tools," terms that never mention AI agents. This mirrors the earlier Perplexity fight almost exactly. The consistent logic: any third-party agent on Amazon's commerce surface is an unauthorized crawler unless Amazon approves it, and the approved substitute already exists in the form of Rufus, Amazon's own shopping assistant inside its app. Amazon is not rejecting agentic commerce. It is rejecting agentic commerce it does not own.

Google Is Not Being Open, It Is Being a Landlord

The tempting read is that Google chose openness where Amazon chose walls. That misreads Home MCP. Access is gated behind Premium Advanced at $20/month, limited to the US, and mediated through OAuth. Google is not giving the smart home away; it is converting Nest cameras, doorbells, thermostats, and Matter devices into a metered API business where rival agents are paying customers. Every Claude or ChatGPT session that checks a camera event history validates Google's position as the home's system of record and generates subscription revenue. The dichotomy that matters is block-and-substitute versus meter-and-charge, and both are monetization strategies for the same asset: a chokepoint others' agents cannot route around.

The Playbook Choice Follows the Margin Structure

Why did Amazon block and Google meter? Follow each company's economics. Amazon's retail engine runs on transaction economics and advertising placement; the company posted $200.6B in Q2 2026 revenue, up 19.6% year over year, with operating margin at 13.7%, up from 11.4% a year earlier. A third-party agent that shops on a user's behalf bypasses search placement, sponsored listings, and Rufus-driven recommendations, threatening the highest-margin layers of that stack. Blocking defends margin. Google's smart home, by contrast, is a hardware installed base with weak standalone monetization; charging rival agents $20/month per user for access converts a cost center into recurring revenue. Neither decision is ideological. Each firm chose the playbook that protects or creates the most margin on its specific surface.

Agent Identity Becomes the Next Picks-and-Shovels Layer

Amazon's core stated grievance was not that Muse existed but that it "did not identify itself." Google's gate runs on OAuth. Both playbooks, blocking and metering, depend on the same missing primitive: a reliable way for an agent to prove who it is, whose behalf it acts on, and what it is authorized to do. Today that layer is improvised from session cookies, terms-of-service language, and pop-up warnings. Whoever standardizes agent identity and authorization (credential delegation, scoped permissions, single-use payment instruments like the Link by Stripe cards Muse already uses) sells to every side of this fight simultaneously. In a war over access, the verification infrastructure is the only neutral ground, and neutral ground in platform wars historically compounds.

Implications

Amazon (AMZN)

Amazon has committed to block-and-substitute: third-party agents are unauthorized by default, Rufus is the sanctioned alternative, and Conditions of Use language written for scrapers is being stretched to cover agents. The strategy defends commerce and advertising margin at a moment when margins are expanding (operating margin rose from 11.4% in Q2 2025 to 13.7% in Q2 2026), but it carries a cost: Amazon is betting users will accept its agent inside its walls rather than bring their own. If personal agents become the default shopping interface, Amazon's block converts from margin defense into demand friction. The Perplexity precedent suggests Amazon will keep enforcing, so the pattern is now legible to every agent builder and regulator watching.

Google (Alphabet)

Google has positioned the smart home as a tollbooth rather than a fortress. Home MCP turns Nest and Matter devices into a paid API surface, with rival agents subsidizing the installed base at $20/month through Premium Advanced. The move also hedges Google's own agent ambitions: Antigravity gets first-party access while competitors pay for the same door. The risk is inversion, since agents that mediate the home could eventually own the user relationship Google currently holds. For now, Google collects revenue either way, which is the stronger structural position on a surface it does not need to defend for margin.

Meta

Muse's Amazon block exposes Meta's structural weakness in the agent economy: it owns capable agents and massive distribution but almost none of the transactional surfaces those agents need. The Stripe single-use card integration shows Meta can solve payments without a platform's blessing, but shopping requires the storefront's cooperation, and Amazon declined. Meta's counter-leverage is WhatsApp and Messenger, chokepoints of its own. Whether Meta opens those surfaces to rival agents, meters them, or blocks them will reveal whether it seeks reciprocity or retaliation, and it is the single most important follow-on decision to watch.

Apple

Apple was absent from both announcements, but iMessage is arguably the most valuable unpriced agent chokepoint in consumer tech. Every playbook question raised by Amazon and Google now lands on Cupertino: block third-party agents from iMessage and substitute Siri, or meter access the way Google metered the home. Apple's history of platform control and services monetization suggests it can run either playbook, and its choice will set the terms for messaging as an agent surface.

Stripe and Identity Infrastructure Providers

Muse's use of Link by Stripe single-use virtual cards demonstrates that payment credentials can already be delegated to agents safely. The unresolved layers are identity and authorization: Amazon blocked Muse partly for failing to identify itself, and Google's gate depends on OAuth. Providers who extend delegation from payments to identity (verifiable agent credentials, scoped permissions, auditable authorization) sit in the picks-and-shovels position of this conflict, selling required infrastructure to blockers, meterers, and agent builders alike.

Conclusion

Two hyperscalers answered the same question in opposite ways within two weeks, and together they defined the agent economy's actual battleground. Amazon proved that agent capability is worthless without permission by neutralizing Muse with a pop-up; Google proved that permission itself is a product by selling it at $20/month. The findings collectively dismantle the assumption that better models win this market: value is accruing to the owners of chokepoints (commerce, payments, messaging, the home) and to the identity layer that governs who passes through them. Block-and-substitute and meter-and-charge are now the two live playbooks, and every platform owner from Apple to Meta must pick one. Access, not intelligence, is the most contested asset in tech.