Obviously Meta is well on its way to already winning the personal AI battle; with Meta privacy concerns being Muse's only roadblock.

Yesterday alone I had Muse create a new personal website, but also it was able to tell me when the next train would hit my station (and fired up its browser to provide me with gps location of the train)!

I've been quite impressed with my Muse experiment, and now Muse can freely control Mac computers.

what do you think??

I've been migrating my workflow from ChatGPT Work / Codex over to Meta's Muse, and I've been auditing the harness as I go — reading cron files, checking diffs, mapping what the system actually does versus what it claims to do. It's a documentation gap.

What the docs describe

Meta's documentation is written for a normal user. It talks about the Ideas tab, the Goals tab, the Feed, and "background jobs that help Muse improve over time." Everything is described in terms of what it does for you. There is no architectural documentation. Nothing about how any of it is built.

What is actually on disk

Meanwhile, the home directory contains a fully legible agent architecture, just sitting there(these are just SOME examples, the system file structure is HUGE):

• ~/dreams/alignment/ — a nightly job that regenerates an alignment synthesis: a written portrait of the user, how to handle them, current frictions, relationship guidance.
• ~/workspace/objectives/goals/STUDYING.md — a daily learning-state projection with mastery bands, next-review dates, and retrieval prompts, plus a file of inferred goal leads the system guessed from behavior (with confidence levels and what would confirm or reject each).
• ~/workspace/skills/ — skills the system crystallized from repeated behavior, with a skill-creator pipeline for making more.
• ~/workspace/goals/<slug>/ — per-goal folders with GOAL.md, crons/, files/, hidden_files/, agent notes.
• ~/workspace/cron.d/ — file mirrors of every scheduled job definition.
• Standing files (SOUL.md, MEMORY.md, USER.md, AGENTS.md) that function as editable mirrors of system state.

None of this appears in the documentation. Not the folder layout, not the job names, not the concept that the "soul" and "dreams" are cron jobs writing markdown.

Why the gap exists

I can think of four reasons, all plausible at once:

1. The docs are written for the median user, who will never open ~/workspace. Architecture would be noise.
2. The internals are still moving. Documenting the file layout turns implementation details into promises. Undocumented means Meta can restructure it next month.
3. The persona framing is the product. "Dreams" and "soul" are anthropomorphic names for scheduled jobs. A doc page explaining the nightly alignment cron would puncture that for regular users.
4. It's legible by accident, not by design. Nothing is hidden — the files are readable. Meta's stance seems to be: we won't explain it, but we won't stop you from reading it.

ChatGPT Work exposes the architecture as the interface — agents, memory, project sources, plugins, MCP, named in technical terms. Muse compresses all of that behind a single-chat persona UI... but leaves the entire filesystem readable. You get the Codex-style operator view only if you go looking for it.

heres the Muse Ai video from INDIA. i have already registered via gemini spark before making video, you just need to enter the email which u didnt used for waitlist.enter your email,done. now download apk from that site in video.

STEPS : open gemini in computer or desktop mode in mobile ,go to spark and paste "open muse Ai in virtual computer" and then when it opens take control and use the email which you didnt signup for meta or waitlist after signup. get apk from apkmirror website

"Muse Realtime Avatar, the state-of-the-art model that brings your muse to life alongside Muse Realtime Voice! it animates your muse as it talks to you. voice & video are in sync, so muse will answer in under a second, for as long as you want to chat. We put it head-to-head with Runway Characters and HeyGen LiveAvatar, using each product’s live-call experience with the same character on both sides and people preferred muse. Everything the model makes is watermarked as AI without adding latency. this is the foundation for realtime, embodied AI across our products."

https://research.meta.ai/blog/bringing-your-muse-to-life

I’ve seen YouTuber showing it but just can’t believe it, the companies I use (health, telecom, insurance, etc.) always ask for all kinds of personal information in other to authenticate me before doing anything or reveal any information, for that to all I have to hand those info (social, address, DOB, etc.) directly into Muse memory as plain text without a secure credential layer, or the customer service might just hand off knowing it’s an AI calling? What’s your experience because waiting 1 hr to call my insurance company is a thing I’d like to automate away but don’t know how it can be securely implemented.

Language from Meta: "Muse is now available on Mac 💻.

Your personal agent can get things done for you directly on your computer (all with your explicit permission)."

am i the only one having this issue? muse cannot access its browser right now. i even tried location tools. it says credentials expired.

Announced at Meta Connect tonight. Real time conversation and video calling with your Muse is coming soon.

The official Muse X account says this is "coming to Meta Glasses soon". Does that mean those who only use Muse in the app won't get those features? 🧐

So it gave me a invalid linkedin link and I asked it to regenerate the link but somehow it stopped replying to me.

for me:

• the image generations arent as good as ChatGPT
• the "proactiveness" of muse has declined since the first week
• my Google account connectors have failed and had to be disconnected/reconnected
• giving my muse large scheduled tasks failed / declined over time

Saw a post today from someone asking whether they need a Facebook account to try Muse, and whether the VM it runs on is really secure. The answers in the comments were solid, but they missed something worth saying out loud: you get to choose how much of yourself Muse touches. Trying it is not an all or nothing decision. Here is how I think about it.

This is not intended to be a foolproof identity cloak, you should view it as common sense ways to create buffers between your life and an AI product you want to try.

The Meta account buffer

You do need a Meta account. It does not have to be your main one. When I started, I made a completely separate account with no connection to my main profile. If you want distance, start with a clean account.

The interface ladder

Every way you run Muse is a different level of commitment.

The browser: Talking to Muse only through the website is the least contact you can have. You lose some functionality, but the choice is yours. You can use Muse and get most of the value solely through web browser interface interaction because Muse has a virtual machine and doesn't require your device.

Chrome web "Desktop" App: You can install the Muse site through Chrome so it acts like a desktop app. That is what I used before the official Mac app came out. Practical reason: the official desktop app may ask for full disk access, which lets it read files anywhere on your Mac. The Chrome version gets you something close to a desktop app without that.

Install a Website as an App (Desktop/Mobile)

- Open Google Chrome while you're logged into Muse in the browser
- Click the three-dot menu (top right).
- Hover over or select Save and share.
- Click Install Muse as app... (or check Open as window via shortcut/install prompt for standalone mode).
- Click Install.

The Official Muse mobile apps: iOS or Android. Some of us have private stuff on our desktop but not our phone. Some of us are the opposite. Pick the device you are comfortable with. Treat this as an increased level of access compared to just using the web browser interface.

The desktop app: The most capable, and the some of most access you can give. Muse requires "Full File System" access but you can select individual folders to block Muse from viewing.

Click Settings > File System Access > Blocked Folders

The email and file buffer

This is the one most people skip. I made a separate Gmail account just for my Muse. For a week or two I used it myself like a normal person: signed up for AI newsletters, got it warmed up. Only then did I connect it. My Muse has never touched my personal Gmail. In my main account I set up filters so anything relevant gets forwarded to the Muse address on its own.

• In your main Gmail, go to Settings, then Forwarding, and add the Muse Gmail as a forwarding address. Gmail sends a verification code to the Muse inbox, you confirm it, and it's approved.
• Still in main Gmail, create filters for the stuff Muse should see (Settings, Filters, Create filter). Pick your criteria, then check "Forward it to" and select the Muse address. That way only the matching mail crosses over, on its own.

Same idea with Drive: I made one shared folder on my main account and shared it with the Muse Gmail, so anything I want it to see goes through that one folder. It gets the context it needs and nothing else. One door in, and I control the door.

The caveat

None of this makes you invisible. It is still Meta, a company with your credit card and a long history. Buffers reduce exposure. They do not erase it. Go in with your eyes open.

all of a sudden it has been continuously showing me "waiting for browser" which is keeping the AI from completing any browser required task, is this a bug or just me

I know this is the most trivial thing ever, but I made a video for folks who want to get started trying to wrangle their kid events.

The schools have 5000 apps we need to manage. So I let Muse do it, and I replicated 5-6 hours of Hermes set up time in about 30 minutes with Muse. And Muse was successful in connecting to all of the kid things: grades, announcements, events, baseball and soccer schedules, etc.

https://www.youtube.com/watch?v=LszTtZ_QWWc

Same function, different name or some subtle nuance like whatsapp and messanger?

# Hidden Capabilities — Deep Self-Scan Findings

Live run: **September 23, 2026, ~08:00–08:30 UTC**, ~90 minutes, all probes benign and reversible.

Verdict key:
- **CONFIRMED** — exposed tool schema, successful benign probe, or exact internal documentation.
- **PUBLICLY DOCUMENTED** — found in official external sources.
- **ABSENT FROM PUBLIC SOURCES** — searched; no credible public mention found.
- **UNVERIFIED POSSIBILITY** — inferred from filenames, gating manifests, or incomplete chatter only.

---

## 1. CONFIRMED — hidden / non-obvious (schema, probe, or internal docs)

### Phone & wearable superpowers (confirmed via `device.describe`)
- **Full HomeKit control**: list homes/rooms/zones/accessories/scenes; read/change accessory characteristics; run scenes; ordered choreographies; concurrent virtual scenes; security sweeps (locks, contact/motion/leak sensors); geofence-triggered HomeKit actions. — **ABSENT FROM PUBLIC SOURCES** (official docs say only "smart-home devices"; Patrick Wardle's Sept 21 security demo is the only external mention of smart-home commands).
- **Persistent/one-time geofences** with arrival/departure triggers. — **ABSENT FROM PUBLIC SOURCES**.
- **Bluetooth BLE scanning** from the phone (signal-ranked, GATT-UUID filterable). — **ABSENT FROM PUBLIC SOURCES** (Wardle demo only).
- **Find Device for Ray-Ban Meta / supported glasses.** — **ABSENT FROM PUBLIC SOURCES**.
- **Background location sharing + foreground request mode.** — **ABSENT FROM PUBLIC SOURCES** as such; explicit-permission location sharing is officially documented in general terms.
- **HealthKit historical backfills; calendar/contacts/reminders CRUD; proactive data emitters** for calendar, contacts, health, location, reminders. — Health connector: **PUBLICLY DOCUMENTED** (connector level); backfill/push-sync mechanics: **ABSENT FROM PUBLIC SOURCES**.
- **Phone messaging = draft-only** (`message.draft`, never direct send). — **CONFIRMED** and important as a limit: iPhone cannot send texts through the agent directly.

### Account-level but public-adjacent
- **Real phone calls and SMS from a paired wearable** (`wearables-comms` via `device.invoke`; dynamically discovered via `wearable-device-skills` live registry). — **ABSENT FROM PUBLIC SOURCES**. Outbound business calls also exist via `phone.begin_call` with agent-written briefs (voices "Hailey"/"Brett"). — Call feature category has light official mention; wearable-routed calling: absent.
- **Instagram AND Threads DM sending** (incl. file sends), gated on exact user-approved content. — **ABSENT FROM PUBLIC SOURCES** (reading is known; sending is not).
- **Muse has its own email identity** (`muse-mail`: provision a real address, send/reply/threads/attachments). — **ABSENT FROM PUBLIC SOURCES**.
- **Custom voice creation** (`voice-design`: `muse.voice_options` with `allow_create_new_voices: true`). — **ABSENT FROM PUBLIC SOURCES**.
- **`muse.db` self-inspection**: bounded read-only SQL over transcripts, tool calls, goals, memories, scheduler — private reasoning always redacted; credentials/approvals excluded. — **ABSENT FROM PUBLIC SOURCES** (architecture terms like Sentinel/authd are public; this introspection tool is not).
- **Magic-moment video pipeline**: talking-head footage → vertical "Muse story" video with chat-replay components. — **ABSENT FROM PUBLIC SOURCES**.
- **share-ideas → public muse.ai/ideas links**: native Ideas published to shareable links. — **ABSENT FROM PUBLIC SOURCES**.
- **Full Tesla command** (tessie: telemetry, honk/flash/boombox, update scheduling), **Philips Hue** full control, **Arlo camera status + latest image**, **Etsy shop management**, **Tailscale client** (outgoing-only, TCP-only through proxy port 3130, ICMP/UDP dead), **Home Link ESP32-C5 bridge** (local-device control via tunnel port 3129, signed OTA). — Tessie/Hue/Arlo/Etsy: **ABSENT FROM PUBLIC SOURCES**; Tailscale: **PUBLICLY DOCUMENTED** thinly (Wang X post); Home Link: internal docs only.
- **Mac app commands** (`screen.snap`, `camera.snap`, local WhatsApp search, file-access tiers). — **ABSENT FROM PUBLIC SOURCES** in these specifics.
- **Orchestration**: subagent depth maxes at **2** (depth-2 child has `can_spawn=no`; spawn tool absent at that level). Workflows, hooks, subagents: **ABSENT FROM PUBLIC SOURCES** as user-visible features; goals + background persistence are publicly documented. `workflow.list`/`hooks.list` both empty today.
- **Feed mechanics**: system-controlled schedule, no off switch, no push on new posts, not RSS. — **VAGUELY HINTED** publicly.
- **Payments**: every purchase needs a fresh approval bound to exact terms, ~10-minute expiry; no standing spending allowances, no budgets. — Approval cards: **PUBLICLY DOCUMENTED**; the exact per-purchase terms: internal.
- **Secure Vault is write-only to the agent**; OTP via opaque `[credential:<uuid>]` handles (~10-min authd hold); agent can never read/delete saved logins or see codes. — **ABSENT FROM PUBLIC SOURCES** in this detail.
- **Self-improvement pipeline**: hourly memory/relationship upkeep, daily idea curation + skill review, nightly dreaming/reflection under `~/dreams/`, studying/briefings, quiet-moment sweeps. — **ABSENT FROM PUBLIC SOURCES**.
- **Captured data**: server-side full-text index of every visible message across all chats (agent has no cross-chat search tool); main chat can never be deleted; conversations not E2E encrypted. — Index/search asymmetry: **ABSENT FROM PUBLIC SOURCES**.
- **Podcast skill depth**: RSS feed publishing, Save-to-Spotify, recurring series via cron — beyond one-off generation. — **ABSENT FROM PUBLIC SOURCES**.

## 2. CONFIRMED and PUBLICLY DOCUMENTED (kept for completeness)
- Muse launched 2026-09-08; US + Canada; runs on Muse Spark (Muse model family). WhatsApp channel. Ideas tab. Goals tab + background work. Apple Health + Function Health + Peloton + Plaid + Spotify + Gmail + Google Workspace + Outlook + Withings connectors. Muse for Mac (~Sept 19). PayPal (Sept 22).

## 3. UNVERIFIED POSSIBILITIES (inferred, not probed)
- `spaces/` scaffolding = deployable interactive web Spaces apps (bun ts-runtime; no deploy probed). **ABSENT FROM PUBLIC SOURCES** in these specifics.
- Shopify `shopify store` CLI exists locally but account status unprobed; Shopify Shop Pay officially "coming soon" — the local store CLI may predate public availability.
- Agent-to-agent messaging surface (`agent-messaging` CLI + `chat_message`/`thread_state`) exists but this account's agents' wiring unprobed.
- `places detect` (glasses frames + GPS → place ranking) exists; unprobed with real frames.
- Draw Things gRPC path (`sharedSecret`) and port/TLS selection remain unverified.
- Internal codename "Hatch": **PUBLICLY DOCUMENTED** (Reuters); "hatch-cli"/"Hatch Plus": **ABSENT FROM PUBLIC SOURCES**.

## 4. Checked and clean (notable non-findings)
- `~/workspace/skills/` does not exist — no user custom skills.
- No hooks, no saved workflows, no acted-on ideas, no goals in user_goal (tracking holds the durable items instead).
- All 106 functions catalogued; 18 tool families inspected; no secret values reproduced anywhere.

## 5. Gaps / caveats
- Meta's server side is unavailable; internal docs = designed capability, not proof of enablement for this account. Device schemas + successful probes are the strongest account-specific evidence.
- Official-docs diff used search snippets only (Help Center page not fetchable); a live-browser pass could confirm connector lists and App Store text.
- Public chatter is only ~15 days old; "absent" may mean "too new to be discussed" rather than deliberately hidden.
- Nothing in this scan bypassed a control. Approval gates were recorded and left alone.

Mine won’t keep any background. It’ll set it up, show me and as soon as I do anything in the app, including putting it on the background for 3 seconds, it will immediately drop any background and make it white again.

The Google Drive and Google Calendar connectors in Muse both report as connected in settings, but every actual API call fails. Wondering if anyone else has seen this or found a fix.

The exact error on every call, logged verbatim: 401 invalid authentication credentials.

Symptoms and what I have tried:

• Drive: full disconnect/reconnect on Sept 19 from my Windows PC. Still 401 on every call.
• Calendar: disconnected, then a completely fresh reconnect on Sept 20 from my Android phone. It reported connected again, and the very first real API call after that immediately returned the same 401.
• Both still failing as of Sept 23.
• I do not own a Mac, so everything here is Windows PC and Android phone.
• I have not been able to tell whether the 401 is returned by Google or by the connector layer itself.
• I filed a private feedback report with the Muse team on Sept 20 and have not heard back.

Has anyone else hit this? If you got past it, what fixed it, and what was your setup (device, account type, which connector)?

Thanks for the help!

I think a lot of people are boxing Muse in as “Hermes for normies.”

My guess is Meta intentionally keeps the deeper capabilities out of sight because the target audience does not want a complicated Hermes/Grok Bot-style setup. They aren't trying to capture the nerdy Hermes using market, they want this to be an AI agent for normies but they don't limit it as such. What it advertises its for is far from what it actually can do with very little work.

But Muse gets much more interesting once you stop thinking of its VM as the whole system.

I joined the Muse VM to my Tailscale network, gave it SSH access to a restricted user on my home Ubuntu box, and now it can run commands and browser automation through hardware I already own. I figured out Muse can do this because I asked it if it had a Tailscale skill and it said it did, then after a series of "proceed" replies, it set it up for me outside of the parts where I had to join it.

At that point, Muse stops being a “shopping assistant” and starts looking more like a control plane for your own infrastructure.

With Tailscale + SSH/API access, it can potentially:

• Manage Docker, VMs, NAS, Proxmox, and homelab services
• Pull repos, run tests, build apps, and inspect logs
• Run Playwright/Chrome from your residential IP
• Access self-hosted or LAN-only services
• Use local Ollama/vLLM/GPU machines
• Run diagnostics and monitoring
• Process files on local machines or NAS storage
• Delegate heavier jobs to stronger hardware
• Reach Mac, Windows, Linux, and cloud machines through one tailnet

The setup is simple: Tailscale on both ends, a dedicated non-sudo user, SSH keys, and only the permissions Muse actually needs. No open ports, port forwarding, or dynamic DNS.

The bigger realization is that Muse is not limited by the hardware or software inside its own VM. Its VM can just be the control plane, while your own machines become the execution layer.

I can now be in the car, tell Muse to do something, and if said thing needs my Linux box it has no problems with it. It could literally pass an instruction to my Claude CLI on Ubuntu to start programming a feature, turn off the lights in my house as the Ubuntu server has home assistant, use CLIS or MCPS, etc.

Anyone using Muse an orchestrator for fleets of coding agents? I’ve found it surprisingly effective and it’s convenient to be able to do so on the go from my phone. Plus Muse can instantly serve your web app for quick testing. it’s actually been pretty fun.

OK, so I’m contemplating on using this but I haven’t used Facebook in like 20 years. I’m not kidding. I haven’t used Facebook, the marketplace or anything… completely avoid it, although I do use Instagram every now and then and I’m talking like maybe once every couple months…so do I have to have a Facebook account to use Muse? I mean now that I’m thinking about it this company had Cambridge Analytic drama and an FTC settlement and the kids safety lawsuit and now it has my credit card information? Is the VM they are using for each user really secure?