Hark Pro vs Meta Muse vs Instinct vs Grok Bot
Compare Hark Pro, Meta Muse, Instinct and Grok Bot from their own launch pages and terms: computers, memory, approvals, pricing and connections.
Hark Pro, Meta Muse, Instinct and Grok Bot are personal AI agents built to act for you, not just answer you. Hark Pro is the newest. Hark Labs, the San Jose AI lab founded by Brett Adcock, released it on October 6, 2026 as a free app for iOS, Android and the web, and calls it "the operating system for your life". Each vendor describes connected accounts, work that continues after the conversation and a way to check in before something consequential happens. The useful comparison is how each one reaches your tools, where you talk to it and which controls you can inspect. This is a launch-day comparison built from each vendor's own pages and terms, not a benchmark of task success.
| Question | Hark Pro | Meta Muse | Instinct | Grok Bot |
|---|---|---|---|---|
| Vendor and home | Hark Labs, Inc., at hark.com | Meta, in Muse and WhatsApp | Spear Street Technology, Inc., at instinct.com | Grok Bot, on x.ai and its own apps |
| Where you talk to it | One persistent chat, plus Projects and a Home screen, on iOS, Android and web | Muse app and WhatsApp | Text or call | Grok Bot desktop and iOS apps |
| Where it does the work | Handoff, its own cloud computer, running up to six browsers at once | Muse Secure VM with its own browser | Inside the accounts you authorize as Connected Services | Its own computer |
| Connected services | Google Workspace and Microsoft 365 by API, MCP, and sign-ins to sites through Handoff | Apps a person connects to Muse | Accounts you authorize, with Google Workspace the named set | Tools and accounts you sign into, plus custom MCP servers added in chat |
| Memory | Remembers what you tell it and learns your preferences; ask it to forget anything | Ongoing goals and connected accounts | Built to understand what you are working on and what matters to you | Memory about you, plus saved routines |
| Work while you are away | Scheduled Tasks, plus Action Buttons it suggests before you ask | Continues projects and returns for changes or approval | When you text or call it | Bots work across apps and come back for approval |
| Agents at launch | One assistant in one thread, with no agents to juggle | One personal Muse | One assistant | Multiple Bots that can coordinate as a team |
| Launch access | Free to everyone on iOS, Android and web | US rollout on iOS, Android and web | Invite-only private access | Beta for eligible SuperGrok and Cursor subscribers |
| Pricing | Every feature free with capped usage; Pro² is $20 a month for 2x usage, Pro³ $100 a month for 10x | Most needs free, with subscriptions for more use | No price published | Bot usage is separate from Grok and Cursor plan usage |
| Connecting a WorkerKit fleet | The WorkerKit MCP server, with a manager key | A custom connector to the WorkerKit API | No connector; worker reports go to its email address | The WorkerKit MCP server, added in chat |
Hark's column comes from its launch article, Introducing Hark Pro, its Using Hark Pro guide, its pricing page, its Security page and its Terms of Service. The other columns come from the Meta Muse announcement, Instinct's own site and the Grok Bot announcement. Plans change quickly in a launch month, so check each vendor's current plan page before treating this as a purchase guide.
Where Hark Pro stands out
Hark describes three pieces: its own computer, memory and proactive thinking. The computer is Handoff, which Hark first showed in an August research preview: a cloud computer, operated by Hark, that signs into websites for you and takes actions such as buying things and filling out forms. Hark says Handoff runs up to six browsers at the same time and can sign in to millions of sites. Outside the browser, Hark connects to Google Workspace and Microsoft 365 by API, and reads files, reaches external databases and finds information over MCP, the open Model Context Protocol. Hark says Handoff is "independently verified to be the most performant computer use model in the world", pointing to its top score on the Online-Mind2Web human-evaluation leaderboard. That is the vendor's claim about a benchmark; the sites you actually use are the test that counts.
The interface is the other distinctive choice. Hark Pro looks like a messaging app: one persistent thread, Projects for longer work with their own files, and a Home screen. Home holds Action Buttons, suggestions that finish with a tap, such as "Pay" on a bill or "Cancel" on an unused subscription, which Hark generates from what it sees in your email and learns in conversation. Home also holds Panels, small apps Hark builds on request over services like Strava or Venmo, which you can then ask questions. Specialties are multi-step workflows Hark learns to repeat, and Scheduled Tasks are created by asking. Hark's launch example is a movie ticket that also brings a parking reservation and an entry QR code, paced so each arrives when it is needed.
Logins and payment cards go in Secured by Hark, an encrypted vault. Hark says even it cannot see inside, and a stored password is detokenized only on the page where it is needed. Its launch promises are short: your data is yours, it will never be sold or shared with advertisers, and you can always delete it, from Settings, then Profile, then Delete Account. Hardware comes next. Hark says its first physical product ships in 2027, and an August partnership with AT&T covers future devices that connect without a phone or Wi-Fi.
Where Muse stands out
Muse is the consumer option with the largest channel already in people's pockets. It lives in its own app and in WhatsApp, and Meta says it is rolling out in the US on iOS, Android and web, with most needs free. Meta describes a dedicated Secure VM, connected accounts, ongoing goals and a Sentinel agent that checks outside actions. If WhatsApp is where your day already happens, that is the practical draw. Our Meta Muse and AI workers guide covers connecting a WorkerKit fleet through a custom connector.
Where Instinct stands out
Instinct, from Spear Street Technology, Inc., is the assistant you text or call, with no new app to learn. It is not Instinct Science, the veterinary software company, and not Deep Instinct, the security company. Its homepage describes an assistant that understands what you are working on and what is important to you, connected to email, messaging, screen, audio and location. Access is invite only. Its integrations run outward into accounts you authorize, and since September 9, 2026 each assistant has its own email address. Hark's Terms also describe text messages from Hark carrying replies and task updates, but its launch presents an app first. Instinct publishes no connector or API for third parties, which is why our Instinct guide delivers worker reports to that mailbox rather than describing a connection.
Where Grok Bot stands out
Grok Bot's launch leads with a team: several bots that coordinate across apps, each working on its own computer and coming back when approval is needed. The beta is open to eligible SuperGrok and Cursor subscribers on desktop and iOS, and Bot usage is counted separately from those plans. It accepts custom MCP servers added in chat, which is how our Grok Bot guide connects a WorkerKit fleet. If you want several agents splitting one project, Grok Bot is the one of the four that leads with it. Hark goes deliberately the other way: one thread, and in its own words no "agents" to juggle.
How do approvals and money differ?
All four say they check in before consequential steps, in different language. Hark's homepage says it "checks in with you before doing something that requires your approval", and its Terms set the default: unless you tell it otherwise, Hark asks you to confirm before it spends your money or sends a communication to other people on your behalf. Its Security page adds the part worth reading twice: Hark starts tasks "by defaulting to consent", and over time it remembers the actions you want it to take. Under the Terms you can also grant standing permission for a category of purchases, after which each purchase inside it needs no prompt. So the check-ins are meant to thin out as Hark learns you. You can watch Handoff act as it works, and Hark says there is always a button to cancel an order.
Messages go out in your name, and Hark says it identifies itself as an AI where the law requires it or where a person is likely to read the message. Hark also reserves the right to decline or stop an action you authorized if it judges the action unlawful or harmful. Its Terms describe optional access to your own device and the sessions you are already signed into, and warn that sites attribute those actions to you; Hark is not responsible if a site limits or suspends an account over automated use.
Meta describes the Sentinel check and asking permission before sensitive steps. Grok Bot says bots return when approval is needed. Instinct acts inside the accounts you authorize. None of these descriptions is a security audit, and the same words mean different things in different products. For Hark, its own pages give a concrete list of what to check.
| Area | What Hark says | What to check in your own account |
|---|---|---|
| Spending money | Asks you to confirm unless you gave standing permission for that category, and learns which actions you want it to take | Which kinds of purchase it no longer asks about, and which card is in the vault |
| Messages to other people | Sent in your name; confirmed first by default; labelled as AI where required | Recipients and wording of anything that recurs |
| Scheduled and proactive work | Later, recurring or conditional tasks within your instructions, plus suggested Action Buttons; pause or cancel any time | Which Scheduled Tasks are live, and which suggestions you accepted |
| Stored credentials | Passwords and cards in the Secured by Hark vault; removable, and deleted when the account closes | Which sites it holds a login for, and whether each still needs one |
| Memory | Ask it to forget anything; it may act on what it remembers | Out-of-date facts it could act on |
| Your content | May be used to improve the service; content flagged in safety review is used for training even if you opted out; third-party models keep no data | Your privacy settings, and what you share in the thread |
| Watching it work | You can watch Handoff act, with a button to cancel an order | Whether you watch purchases live or review them afterwards |
Compare one real task before choosing
A launch demo shows what an agent is designed to do. Your own errands show whether it does them with the access and review you need. Pick one ordinary task that takes 15 to 30 minutes and that a person can check. Two good candidates pull in different directions. "Find three restaurants for Thursday that suit this group and book the best one" plays to Handoff. "Prepare a briefing for my next meeting from the invitation, recent email and the project document" plays to connected accounts and memory. Give each agent the same sources and the same definition of done, then record:
- Source access: which messages, documents and sites it used, and anything it reached that it did not need.
- Work completed: how many corrections the result needed before it was usable.
- Approvals: what it asked before acting, and anything it did without asking.
- Continuity: whether it can pick the task up again tomorrow without being re-briefed.
- Cost: which allowance the task drew on, and whether it stays practical at daily volume.
Where an AI worker fits beside a personal agent
A personal agent is at its best on the errand you hand it today. A job that must run the same way every morning, with its own app grants, model, budget and record of each run, is a different shape of work, and that is what a WorkerKit worker is for. The executive daily briefing is ready before you open any chat; the follow-up tracker catches a missed reply without anyone asking a personal agent to search every morning. Review each kit's permissions before installing it.
The boundary matters more as personal agents gain reach. Hark's can include a vault of your logins and cards. A worker holds its own scoped grant, read-only by default, so a standing job never rides on the access you gave a personal agent, and each run closes with a run receipt carrying the result and its settled cost.
Each agent reaches a fleet through a different door: Muse through a custom connector, Grok Bot and Hark through MCP, and Instinct through email delivery. On October 6, 2026 we connected Hark Pro to the WorkerKit fleet as a remote MCP server, with a manager key, and it listed the workers and read the fleet's health on its first calls. Our Hark and AI workers guide and its setup steps cover that connection, including how to keep the key narrow, since Hark asks for it in the chat. Mail is a useful second channel: Hark builds its Action Buttons partly from your email, so a worker that delivers its report to that inbox is reporting where Hark already looks. For a careful first job, our guide to rolling out an AI worker safely starts small and reversible, and our earlier comparison of OpenAI Dots, Meta Muse and Grok Bot covers the agents that came before Hark.
FAQ
Who makes Hark Pro?
Hark Labs, Inc., an AI lab in San Jose, California, founded by Brett Adcock, who also founded the humanoid robot company Figure. Hark is the company and the personal intelligence it is building; Hark Pro is the app, on the web, iOS and Android. Hark raised a $700 million Series A at a $6 billion valuation in May 2026 and plans its first hardware for 2027.
Is Hark Pro free?
Yes, with a cap. Hark says every feature is free in every plan, and upgrading only raises usage: Hark Pro² is $20 a month for twice the free usage and Hark Pro³ is $100 a month for ten times it. Auto top-up refills credits by an amount you choose, up to a ceiling you set. Those are the web prices. The App Store lists the in-app subscriptions higher, at $26.00 for Pro² and $129.99 for Pro³, plus a Pro⁴ at $200.00 that the web pricing page does not show, so compare before subscribing inside the app. Before launch, Brett Adcock said the first 100,000 sign-ups would get a paid plan free; Hark's pricing page does not describe that offer.
Does Hark use its own AI models?
Partly, by its own account. Handoff, the computer use agent, is Hark's own model: Hark started with post-training, has moved into mid-training and plans pre-training for later in 2026. Its Security page says Hark "prioritizes its own models whenever best" and that models from vetted third parties keep zero data. Hark's current Privacy Policy does not name those providers.
Does Hark train on my conversations?
Hark's Terms say it may use your content to improve the service and develop products, and that content reviewed for compliance, safety or misuse is used for training and quality assurance even if you opted out. The third-party AI providers it uses are not permitted to train on your content, and Hark promises never to sell your data or share it with advertisers. Memory is yours to edit: ask Hark to forget something and it will.
Can Hark Pro connect to WorkerKit?
Yes. On October 6, 2026 Hark Pro connected to the WorkerKit fleet as a remote MCP server, with a manager key in the Authorization header, and answered its first fleet calls. Hark asks for that key in the chat, so create one for Hark alone with the narrowest scopes and an expiry; our Hark setup guide walks through it. A worker can also deliver its report to the inbox you connected to Hark, and the fleet stays reachable from any MCP client, the CLI or the API, so the workers run whichever assistant you pick.
Is there a Hark device?
Not yet. Hark says its first physical product will ship in 2027, with Hark Pro as the software foundation that comes first, and its August partnership with AT&T covers future devices that connect without a phone or Wi-Fi. Hark has not shown the device.
Which one should run recurring work?
Hark Pro has Scheduled Tasks and Specialties, Muse continues projects between conversations, and Grok Bot has saved routines. Test the exact recurrence, permissions and review path you need before choosing. A separate worker schedule is the better fit when one job needs its own instruction, access, budget and receipt.
Sources: Introducing Hark Pro, Using Hark Pro, Hark pricing, Security at Hark, Hark Terms of Service, Hark Privacy Policy, Introducing Hark Handoff, Hark and AT&T, Hark Pro on the App Store, Meta Muse, Instinct and Grok Bot.