Meta's Muse is an artificial intelligence (AI) agent built by Meta—an autonomous system that can browse the web, use apps, and carry out tasks on a user's behalf. It launched in September 2026, pitched largely at consumer uses like managing calendars and ordering deliveries. In a first-person post dated October 1, 2026, the developer Scott Cooper writes that those consumer features turned out to be the least interesting part of the product for him.
The article is built around Scott Cooper's first-person account of discovering that Meta's Muse agent is far more useful for web scraping than for the consumer tasks it was designed for. This is the cover graphic from that original post.
Cooper paid $80 for a month of Muse and reports that he "hasn't found much use for the personal assistant part." He orders DoorDash about five times a year, doesn't get many emails, and finds his appointments easy enough to manage already. What he does need is data—from Reddit and YouTube—for a side project called fullsets.fm, which aggregates professionally recorded concerts (festival livestreams, Tiny Desk sessions, and similar), organizes them by artist, and extracts tracklists.
He already had part of a pipeline in place: Luna and Gemini 3.8, run through Gemini Studio, screen individual YouTube videos for quality and content. What was missing was discovery. Cooper can scan a single channel, but good concert footage is scattered across obscure channels—T-Mobile Europe, for instance—that have uploaded a handful of random concerts. Muse handles that end. He can instruct it to scan Reddit and YouTube for a thousand videos matching his criteria, or to build a list of every artist who has ever played at Coachella and then search YouTube for each one in turn. It sits there and runs until the job is done, "punishing those websites with a machine in the cloud that has seemingly endless bandwidth and a decent CPU." He also uses it at the other end of the pipeline, for artist matching and a final approve-or-reject call on each video.
Technically, what Cooper is running is a managed virtual machine with a copy of Chrome, driven by the Muse Spark 1.3 model. The agent operates through a real browser—part of what, in the author's experience, lets it present itself as an ordinary visitor rather than a scripted scraper. The author does not describe how Muse handles captchas or other challenge pages, and the supplied sources don't establish whether it can pass them on its own or needs a person to intervene. If one process is too slow, he can spin up more. He reports using a few billion tokens a week on this work—tokens being the model's unit of processed text—and says that a lot of the time the usage doesn't seem to get counted at all.
Cooper is candid about the model's limits. Muse Spark 1.3 is, in his words, "pretty mediocre at writing," and he would "never use it to code." What he finds genuinely useful is what he calls a "Gemini Flash-style workload"—high-volume, repetitive search and filtering, where tirelessness and cheapness matter more than judgment. That distinction matters, because the scraping use case doesn't require a brilliant model; it requires a model that won't stop.
The same qualities that make Muse attractive to users make it a problem for the sites it visits. Cooper notes that Amazon "blocked it already," and predicts that "we'll likely see more websites block Muse unless Meta prevents abuse." Reporting from The Verge confirms the block in concrete terms. Amazon began displaying a popup for Muse users stating that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use." According to GeekWire, cited by The Verge, Meta had not notified Amazon that Muse would access its store, and Amazon raised privacy and security concerns: the agent failed to identify itself while browsing, and it appeared to be capturing customer credentials. An Amazon spokesperson said third-party apps that purchase on behalf of customers "should operate openly and respect service provider decisions about whether or to participate."
This is not an isolated incident. The Verge reports that Amazon sued Perplexity in November of the preceding year over a similar attempt to keep its platform out of an AI search experience, with a judge siding with Perplexity in August. Amazon has also made its confirmation emails sparser since July, omitting item names and product images to limit what external AI services can mine. The Muse block fits a pattern Amazon has been building for some time.
The practical consequence is a cat-and-mouse relationship between agent builders and website operators. Cooper worries about what happens when someone points Muse at his own websites: the agents don't sleep, they can impersonate real users convincingly, and scaling up is simply a matter of spinning up more processes. The cost lands on the operators whose bandwidth and CPU absorb the traffic, and whose bot defenses get continuously re-tested. Whether Muse's traffic comes from a narrow set of IP ranges that sites can block wholesale, and whether Meta will shift to residential proxies to sustain access, are questions the supplied sources don't answer. Cooper himself notes rough edges that temper the picture: jobs can freeze mid-run, communication with the model can be frustrating, and the interface doesn't handle multiple browser jobs well.
So what does the evidence actually establish? Cooper's post establishes that Muse worked well for his specific scraping workload, at $80 a month, with the caveats he lists. The Verge reporting establishes that Amazon blocked the agent and the reasons it gave. What neither source establishes is how far the blocks will spread, whether Muse's infrastructure will adapt, or whether the model will improve enough to change Cooper's view of its writing ability. The post includes a referral code—an element worth noticing, even if it doesn't bear on the agent's capabilities. For now, the most defensible reading is that Muse is, for the right workload, a surprisingly effective scraping tool—and that its effectiveness is exactly what is drawing the resistance.