It is 9:40 on a Tuesday night, and you have a doctor's appointment in the morning. Your phone already knows how you slept for the last three weeks, how many steps you took, and what your resting heart rate did after that flu in October. Somewhere in the Muse app sits a button that says, in effect, hand all of that to your agent and it will write you a one-page brief for the visit, five questions worth asking in plain language, ready to print, and all it asks in return is access to the most intimate record your phone keeps.
That brief is real, and it is the best reason to connect your health data to an agent, because a doctor who sees three weeks of sleep and a resting heart rate trend in the first minute of an appointment asks better questions than a doctor who spends that minute asking when you last slept well. It is also the moment to stop and read what the button actually does, because a health connection is not like connecting Spotify. Music taste is embarrassing. Heart rate is intimate, permanent in a way a playlist never is, and the chain of custody between your wrist and Meta's cloud has more links in it than the connection screen admits.
Muse has offered this connection since launch day, September 8, 2026. Apple Health was live in Muse from the start, on iPhone only, with access managed in your phone's settings rather than in Muse's web connector list, while on Android the equivalent is Health Connect. Independent guides describe the setup as taking about a minute: open Muse, find Apple Health, approve the categories you want in iOS Settings, and the agent can read your steps, workouts, sleep, and other metrics from the Health app.
What it can do with that access is genuinely useful and narrowly bounded. Muse can summarize your sleep, flag the nights under six hours, spot a pattern in your activity, and draft the questions for your appointment, but it cannot diagnose anything. It cannot tell you whether a reading is safe, and it cannot replace a clinician. One independent guide puts it bluntly: Muse is not a doctor, and this connector does not make it one. Meta has published no detailed list of exactly which data types Muse can read, and nothing at all about whether it can write data back to the Health app, and no independent auditor has published a test of which categories the connection actually requests when a user taps through the flow in the current app version. Until Meta says otherwise, treat the connection as read-only, and treat the silence as a gap, not a guarantee. Silence is not a guarantee. Read the list.
Your phone guards this data better than your agent does
Start with the part of the chain that works well, because it deserves credit. Apple built HealthKit around a permission model that is unusually careful. Every data type is a separate grant, and reading is separate from writing. You can let an app read your step count while refusing it your blood glucose, and you can let it write a workout without letting it read your sleep. If you refuse a read category, the app is not even told you refused; its query returns the same empty result an empty database would return. Health data on the device is encrypted while the phone is locked, and with two-factor authentication turned on, it is end-to-end encrypted in iCloud, which means Apple itself cannot read it.
Apple also sets rules no app can contract around, and they are written into the App Store Review Guidelines, which means an app that breaks them loses its distribution, a penalty with real teeth in a way a privacy policy's promises do not have. HealthKit data may not be used for advertising, may not be sold to data brokers or information resellers, and may not be disclosed to a third party without your express permission. Every app that touches HealthKit must publish a privacy policy describing exactly how it uses the data. Revoking is one screen: Settings, Health, Data Access and Devices, switch off the category, and it takes effect immediately.
Google's Health Connect, the Android equivalent, works on the same principle with a longer list. Its permission reference names more than 42 record types, each with its own read and write permission: steps, heart rate, sleep, blood pressure, blood glucose, body fat, menstruation, sexual activity, and, in the medical section, conditions, laboratory results, medications, and personal details. Two meta-permissions sit above the list: one grants the entire history of every type, and another grants background reads of every type. Both are separate gates, and both are worth noticing, because a health app that can read your history in the background is an app that can build a timeline of your body without ever asking twice.
Then the data leaves your phone
All of that careful design governs what an app can read from the store on your phone. It says nothing about what happens after Muse reads it, because at that point the data lives in a cloud virtual machine Meta calls the Muse Secure VM, and a different set of rules applies.
Meta's own launch and safety materials, both dated September 8, 2026, describe a real architecture. Each person's Muse runs on a dedicated virtual machine. A separate system called Sentinel approves or blocks anything the agent sends to the internet, and sensitive actions need your approval. Muse cannot see your passwords or payment methods, and Meta says it does not share your conversations or the data in your VM with its advertising systems. Those are specific promises, and they are worth more than the vague reassurances most companies offer.
Read the same materials one paragraph further and three more facts appear. First, the current Secure VM "does not prevent Meta from accessing data when necessary to support, secure, or operate the service." Meta restricts its own staff access through operational policy, not through cryptography. Second, your conversations, tool calls, and subagent handoffs, collectively called trajectories, are sanitized and used to train new model checkpoints by default, unless you find the switch buried in Settings under Data controls and turn it off before your first health question ever leaves the app. Third, the version that would close the door entirely, Muse Confidential VM, encrypted with a key only you hold so even Meta cannot read it, was announced for later in 2026 and was still not live as of November 3, according to every independent check of Meta's help pages.
One more line from Meta's privacy policy, which took effect September 17, 2026, belongs in any honest account. After you delete something, Muse "may still 'remember' information it learned from what you deleted." Deleting a health summary from your chat does not delete the lesson the agent drew from it.
Count the decisions in one tap
Now put the two halves together, because the gap between them is the whole story. On your phone, health access is dozens of separate decisions: each data type, each direction, each meta-permission, while in Muse, connecting is one tap. Meta has never published which HealthKit categories Muse requests, so the reader cannot know whether that tap stands for steps and sleep, or steps and sleep and heart rate and blood pressure and everything else the store holds. One tap, up to forty-odd permissions, no itemized receipt.
That is the first number that matters. Here is the second: the cost of a bad grant is not ongoing access, because you can cut that in seconds on either platform; it is the copy already read. Revoking a Health Connect permission stops future access; it does not erase data the app already copied, and on Meta's side deletion does not even guarantee forgetting. A health summary Muse read in November can be sanitized into a training trajectory, summarized into a memory file, and quoted back to you in February after you revoked the connection in December. The decision that matters is the first grant, not the revoke, and it is the one the interface makes cheapest.
There is a third number, and it is the one nobody's toggle covers. Health talk is rarely about one person. You tell Muse about your mother's blood pressure, your child's inhaler, your partner's sleep apnea. In October, TIME reported that Muse updates a running dossier every hour on the user and on every person mentioned, including people who never signed up for Muse, and Meta did not dispute the findings, which means a file on your mother can exist on Meta's computers because you once mentioned her blood pressure in a chat, and she has no settings screen anywhere that lets her see it or delete it. Apple's per-type permissions protect your data. Nothing in the chain protects the people you talk about, and no settings screen lets your mother opt out of a file your agent keeps on her.
The grant ladder
None of this means you should refuse the connection. It means you should climb it one rung at a time, in this order, with the settings paths written down before you start. Write them down.
First, before connecting anything, turn off training. Open Muse, go to Settings, then Data controls, and switch off "Help improve our AI models." Meta's policy says the setting is on when you first use Muse, and switching it off covers earlier interactions too. Do this first, because it is the only step that limits what Meta does with data after Muse reads it, while every other step in this ladder only controls what Muse can read in the first place, which is a different problem from what happens to the copy once it exists.
Second, connect with steps and sleep only, and read-only. In iOS Settings, under Health, Data Access and Devices, grant exactly those two categories and nothing else, then use the connection for a month. Read what Muse actually summarizes, and check it against the Health app before you trust it.
Third, add heart rate only if the summaries earn it. Heart rate is where the summaries stop being a calendar and start being a body, and it is the category where a misread pattern causes the most anxiety for the least benefit.
Fourth, never grant write access. Muse has no published reason to write to your Health app, and a write grant is a second, separate door you do not need open. Never grant the history or background-read permissions on Android either. A doctor-visit brief needs the last few weeks, not your entire medical timeline, and it does not need to read while you sleep.
Fifth, hold off on lab results. HealthEx can hand Muse your actual medical records, and the doctor-visit guides recommend it, but lab results are the most sensitive data in the chain, and the Confidential VM that would keep even Meta from reading them does not exist yet. Wait for it.
The case for connecting, at full strength
Fairness demands the strongest version of the other side, and it is strong. Apple and Google built these permission systems precisely so you can grant steps and sleep while withholding glucose, medications, and cycle data, and Apple bars advertising use of HealthKit data outright, which is a harder rule than most health apps live under. Meta's promises are specific and testable: no ad sharing, credentials out of the model's reach, Sentinel approval on outbound actions, a full audit trail, a training opt-out that covers past interactions. A reader who follows the ladder above gets a genuinely useful doctor-visit brief from data their phone already collects, and the agent never sees anything the phone did not already hold. The real failure mode is not the architecture. It is the user tapping "Allow all" without reading the list, which no system can fix. That argument deserves its full weight, and the ladder is built to honor it: connect, but connect narrowly, and only after the training switch is off.
One more piece of history belongs in the record before anyone decides. In 2022, The Markup found Meta's tracking code on 33 of Newsweek's top 100 hospitals' websites, and inside the password-protected patient portals of seven health systems, where it sent Facebook details including the names of patients' medications and the medical condition behind an appointment booking. Meta's response was that a filter was supposed to strip that information out, a defense that concedes the code was there in the first place and asks you to trust a filter you cannot inspect to catch what a tracker you never approved was already collecting. That was a different product, four years ago, and the company has since built the most detailed safety architecture of any consumer agent. It is also the same company, and the same promise, "we will not use your health data this way," made by the same institution that once had to explain why its code was inside a patient portal. A promise is not a verification. Check it yourself. The permission screens are verification, and they are the part of this story you can check yourself, tonight, before you tap Connect.
What I could not verify
Honesty requires the gaps stated plainly. No hands-on test was possible for this article: no Muse account was used, so I could not confirm which HealthKit categories Muse actually requests, whether it asks for write access, or what the connection screen discloses, and no reader should take the read-only claim in this article as settled fact until Meta publishes the list or an independent tester records the flow. Meta has published no list of the exact data types Muse reads, so the read-only characterization is an inference from silence, flagged as such. Confidential VM status was last checked November 3, 2026, and no update was available after that date. No independent audit exists of whether Muse's health summaries are accurate, or of what "sanitized" removes from a training trajectory in practice. Whether the Android Health Connect path is live in Muse, or only Apple Health, is less firmly documented than the iPhone side. Treat this article as an audit of the published rules and the published gaps, not as a hands-on test.
Connect narrowly, or not at all
Your phone already holds the data. Your agent wants to read it, and the brief it can write is worth having before a morning appointment. The price is not money. It is a copy of your body, sitting in a cloud computer that Meta can access when necessary, summarized into a memory that deletion does not guarantee forgetting, and used to train models unless you flip one switch first. That is a price worth paying for a good doctor-visit brief, if you pay it with your eyes open, one category at a time, starting with steps and sleep, and stopping the moment the summaries stop earning their keep.
Tap Connect tonight if you want, but turn off training first, grant two categories, and read the list before you grant the rest, because the only grant you cannot take back is the first one. Your heart rate is the most honest number your phone owns. Hand it over like it matters, because it does.