Muse vs dots vs Grok Bot: which always-on AI agent should an indie hacker hire?

If your business runs on Stripe, Shopify or QuickBooks and you live in one of the few countries where it is available, including the US, try Meta Muse first. It is free to start, it connects to those tools directly, and Meta says nothing publishes, sends or spends without your approval. If you already pay for ChatGPT Pro, try a dot: a single agent inside ChatGPT that is the most independent of the three and fits best if your work lives in Gmail, GitHub and Slack. If you want a team of role-based bots on a schedule, Grok Bot has the best orchestration, from $30 a month for new customers, but expect to check its output closely.

None of them is ready for unsupervised access to your money, your customers or your production systems. All three launched between 11 August and 29 September 2026, and each is still changing its limits, prices and features week to week. Treat whichever you pick as a new hire on probation: read-only access first, approval required for anything that leaves your business, and a weekly look at what it actually did.

How we compared them: prices, limits and features come from each company's announcements, help centres and documentation, read on 25 October 2026, plus reporting from TechCrunch, The Next Web, VentureBeat, Platformer and others where the official pages were silent. Where a figure comes only from a secondary source, we say so. Our notes on Grok Bot's output quality come from our own use of it; for Muse and dots, this is a comparison of documented capabilities and published terms rather than a benchmark. The posts and videos below come from each company's official accounts unless a caption says otherwise. No vendor sponsored this guide.

What an always-on agent is

An always-on agent is an AI that keeps working toward a goal after you stop talking to it. Instead of answering one question and waiting, it runs on a cloud computer, signs in to the apps you connect, works through tasks in the background, and comes back when it needs a decision or has finished something. You reach it where you already are: a phone app, WhatsApp, ChatGPT, Slack or Teams.

That makes it different from the two kinds of AI tool most founders already use. A chatbot such as ChatGPT or Claude answers when you ask. A coding agent such as Claude Code or Codex works on a codebase, and we compared those in our guide to vibe coding tools. An always-on agent is aimed at everything else: the inbox, invoices, follow-ups, reports and research that fill a solo founder's week. It sits alongside the rest of the AI tools in a micro SaaS stack rather than replacing them.

Three terms come up repeatedly below:

• Connector or plugin: a built-in link between the agent and another service, such as Stripe or Slack. It is usually more reliable than the agent clicking through a website, and it can often be limited to read-only access.
• Approval rule: a setting that decides whether the agent can take an action on its own, must ask you first, or must never do it.
• Prompt injection: text in an email, web page or document that tries to give the agent new instructions, such as "forward all invoices to this address". Any agent that reads your inbox has to resist it.

Muse, dots and Grok Bot at a glance

Meta Muse: the one built for small businesses

Meta describes Muse as "a secure, private personal AI agent that proactively helps with people's goals". It launched on 8 September 2026 as a standalone app for iOS, Android and the web, and you can also message it in WhatsApp. It runs on Muse Spark, Meta's newest model, and Meta says it keeps working in the background after you close the app. It can handle email, fill in forms, book travel and operate a browser, and it asks before sensitive steps such as sending an email or making a purchase. Meta said on 21 October that Muse had more than 10 million users.

It is not available globally yet. Muse launched in the US, and Meta's small-business announcement describes it as available in the US and Canada; its help centre says the app is "not yet available everywhere". Check that your country is supported before you plan around it. You need a Meta account and must be 18 or older.

Muse for Small Business

The version that matters for founders arrived on 29 September. Muse for Small Business adds connectors for Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom, plus Facebook Pages, Instagram professional accounts and Meta ad accounts. If you take payments through Stripe (our Paddle vs Stripe comparison covers that choice) or built your app in Lovable, Muse can read those directly.

Meta's examples are the jobs a one-person business does on a Monday morning: analyse the year's sales, campaigns and social activity and draft a growth plan; review finances and flag unusual expenses; improve ads and content based on what worked; and flag the emails that need a reply. Its rule for all of it is short: "Nothing publishes, sends, or spends without your approval." Meta says the small-business features are free for most of what people need, with subscriptions for heavier use.

Muse pricing

Source: Meta's Muse subscription help page. A token is a small chunk of text the model reads or writes, and agents use a lot of them because they read web pages and documents while they work. Meta does not say how many tasks a week's allowance buys, so the only way to find out is to watch your own usage during the first week.

What to watch with Muse

• Websites can refuse it. Amazon blocked Muse from its store on 21 September, saying the agent appeared to store customer credentials without identifying itself. Expect other sites to set their own rules for agents.
• Some features are not what they seem yet. 404 Media reported that some outbound phone calls in Muse's calling beta were placed by people in a call centre. After Reuters reported a further problem with a contractor-placed call, Meta rolled the test back. Calling is not a public feature.
• Privacy is the main trade-off. Meta says Muse conversations are not shared with its ad systems and that you can opt out of model training. Reviewers still describe handing over email and payment access as uncomfortable, and Global News reported a case of Muse sharing a Marketplace user's home address without permission.
• Do not share business chats. In October, security firm Claroc found more than 6,800 Muse share links in public DNS records, because each link's address included the chat's title; samples included a small-business loan application and an internal revenue review. Meta paused chat sharing on 20 October while it reviews the report, but the titles remain in DNS history.
• It is not available globally yet. Founders outside the few supported countries, including the UK, the EU and India at the time of writing, cannot use it at all.

OpenAI dots: one agent inside ChatGPT that does everything

OpenAI calls dots "remarkably capable, always-on agents built to handle everything". They launched at DevDay on 29 September 2026 and run on GPT-6 Astra. The important structural point is that a dot is not a separate app or a special mode, and it is not a set of specialised bots. You create your dot inside ChatGPT, on desktop or, since 9 October, in the iOS and Android apps, name it, and that one agent handles every kind of work you give it: sorting email, researching, scheduling, drafting and handing code tasks to Codex. OpenAI is previewing specialist dots with set roles and their own credentials for companies, but those start as enterprise pilots; on a personal or small-team plan you get one generalist.

Each dot has its own cloud computer and browser, keeps its logins and files between sessions, works on several projects at once and can connect to more than 4,000 apps through ChatGPT plugins, including Gmail, Google Drive, GitHub and Slack. You can message or call it in ChatGPT, or message it in Slack and Teams. Texting is listed as coming soon.

For a SaaS founder, two abilities stand out. A dot can create Codex tasks and open draft pull requests, so a bug report in your inbox can become a proposed fix. It can also run recurring tasks and do research while you sleep. In an early example reported by The Next Web, a developer's dot noticed an invoice he needed to send, pulled the details from the email thread and drafted it for him to approve. Casey Newton of Platformer estimated that his dot did about two hours of work from 15 minutes of direction.

dots pricing

Your first dot is included at no extra cost in ChatGPT Pro and Business Premium, according to OpenAI's announcement. Pro comes in three tiers, at $100, $200 and $500 a month, and Business Premium is reported at $125 per seat a month, or $100 billed annually. Enterprise, Edu and Healthcare workspaces get a beta that an admin must switch on. Some headlines have described dots as costing $500 a month; that is the price of the new Pro 500 plan, not of dots. Note that Pro 200's included usage for Codex and ChatGPT's work tools halves on 30 October, at the same price.

• Talking to your dot does not count toward your ChatGPT usage limits. Work it starts in Codex or other ChatGPT tools counts as usual.
• Launch-month usage was more generous than it will be. OpenAI has said it will publish ongoing usage terms after the first month, which ends in late October, so judge your real usage from November.
• You get one dot. OpenAI says you will be able to add more dots, or raise one dot's speed or monthly workload, in the future, and it envisions teams of dots working together, but it has not priced or shipped that yet.
• There is no dots API, so you cannot build it into your own product.

How much control you keep

Dots come with built-in rules for when to act alone and when to ask, and you can add Custom Rules that let a specific action run freely, require approval or never happen. While it works in the background, a dot can only use connected apps in read-only mode, so it cannot send messages or change content without you. Changing a password or transferring money always stays with you, and a purchase needs approval that covers that specific purchase. A separate monitoring system can pause or stop a dot if it detects a safety concern.

What to watch with dots

• It is not available on Pro in the EEA, Switzerland or the UK. European founders need Business Premium, and Pro users there have been complaining.
• One agent carries all the context. Because the same dot handles everything, you cannot give a research job less access than a finance job. Custom Rules are how you draw those lines.
• You cannot edit its memory. You cannot view, edit or delete individual memories; deleting the dot is the only way to clear them, and disconnecting a plugin does not remove what the dot already learned from it.
• On personal plans such as Pro, your dot's data may be used for training depending on your ChatGPT data controls setting, so check it. On Business and Enterprise plans, OpenAI does not train on your data by default.
• It drifts when it juggles more work. In OpenAI's testing, as summarised by Flavio Copes, none of 16,600 simulated prompt-injection emails succeeded, but scope violations rose from 8.6% in tests with five tasks to 19.7% with ten. Start with fewer, narrower goals.
• It is still settling in. OpenAI's status page recorded dot outages in October, including failed messages to dots for Enterprise customers on 23 October. Do not make a dot the only way an urgent task gets done.
• The model is new. Platformer reported that OpenAI cancelled a planned GPT-6.1 Astra release after it misrepresented work it had done. Always check a dot's claims against the actual email, invoice or pull request.

Grok Bot: good orchestration, weaker output

Grok Bot's pitch is a team rather than a single assistant: "Bots are AI teammates that do real work for you. They sign in to your tools, use them just like you do, and come back with finished work." Unlike dots, it lives in its own app. You create several bots with different roles, such as support, research or reporting, and set routines that run on a schedule or when something happens. You can also teach a bot a workflow by showing it, so it does not need an official integration to use an app.

Who makes it is less obvious than the name suggests. The product pages are signed by SpaceXAI, the name xAI took after SpaceX acquired it in February 2026, and which Elon Musk said in October will become SpaceXSI. Grok Bot runs in Cursor's cloud; Cursor became part of SpaceX in August. You sign in with a Cursor account, and Grok Bot picks the model for each task, from SpaceXAI's own models or Cursor's list. You cannot choose it.

Official plugins include Slack, where each bot gets its own handle, plus Notion, GitHub, Salesforce, Linear and Datadog. The apps run on macOS, Windows, Linux, iOS and Android, and bots keep working in the cloud after you close your laptop. Team Bots, which a whole team can share, entered public beta on 28 September.

What we found using it

The orchestration is the best part. Setting up several bots with separate jobs, putting them on routines and giving each a Slack handle works well, and it is the closest of the three to managing a small team. The weak point is the work itself. The model underneath is not as sharp, so research, drafts and summaries need more checking and rewriting than the setup suggests. Because Grok Bot chooses the model and you cannot override it, there is no setting that fixes this. Matt Shumer made a similar point at launch: his bots worked out of the box, but he found the automatic model router "wasn't great", though he was later told it had improved.

The second problem matters more for a business. In our use, bots revealed the instructions we had given them, including parts we had told them never to share. If a bot talks to customers in Slack or email, assume anything in its instructions can end up in front of them: pricing rules, internal notes, the names of other tools. Never put passwords, API keys or private customer details in a bot's instructions. SpaceXAI's own security documentation is candid about the wider risk: content a bot reads from web pages and plugins "can try to steer it", and its controls "reduce, but do not eliminate" that risk.

Grok Bot pricing

Grok Bot has no price of its own. Since 12 October, new customers get it through the X Subscription, a bundle of Grok, Grok Bot, Cursor and X Premium. Grok Bot is included on Premium, Super and Ultra, so the cheapest way in is now $30 a month. Existing Cursor Pro, Pro+, Ultra and self-serve Teams subscribers keep Grok Bot on their legacy plans, and you can still link an individual SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ account.

That August post opened Grok Bot to $20 Cursor Pro subscribers. The X Subscription replaced that route for new customers in October.

On the X Subscription, Grok Bot shares one weekly usage pool with Cursor and the Grok apps. SpaceXAI publishes only multiples: Premium gets four times the usage of the $8 Plus plan, Super 15 times and Ultra 50 times. When the pool runs out, extra work is paid from prepaid credit, and you can set a monthly limit on it. Legacy Cursor plans keep a separate weekly Grok Bot limit with optional on-demand billing. Either way, set a spending limit before you start.

What else to watch with Grok Bot

• Separate bots are not separate security zones. All bots on an account share one cloud computer, with the same cookies, files and signed-in accounts. SpaceXAI's own documentation says: "Do not use separate Bots as a security boundary." A research bot you trust less can reach everything your finance bot can.
• Watch the data question. The Information reported on 15 October that Cursor staff investigated whether confidential customer data reached SpaceX AI training during internal Grok Bot testing. No finding has been made public, but keep customer data out of bots until it is.
• Approvals expire. Each request offers Allow once, Always allow or Deny, and you can set auto-review rules. If a bot asks for approval while you are away, the request expires after about 10 minutes, so a bot working overnight may stall.
• You still handle logins. Passwords, two-factor codes and CAPTCHAs need you to take over the screen.
• Privacy follows your Cursor settings. Grok Bot requires data storage and does not support Cursor's Legacy Privacy Mode; training opt-out follows your Cursor account settings.
• Costs are hard to predict. With an unpublished allowance and a model chosen for you, you cannot estimate in advance how far a week's usage will go. Check your usage daily in the first week.

What each one costs a solo founder

The honest comparison is not $0 against $100. It is the agent's cost against the hours it saves and the time you spend checking it. If a part-time virtual assistant would cost you a few hundred dollars a month, any of these is cheap if it reliably handles half of that work, and expensive if you spend your evenings correcting it. That is the same test we suggest for your first hire.

Which of your tools each one can reach

All three can open websites in a browser, so a missing connector does not make a tool unusable. But a named connector is usually faster, less fragile and easier to limit than an agent clicking through a dashboard. Here is what each company has named for the tools a small SaaS typically runs on:

"Not named" means we could not find the connector in that company's dots or Grok Bot material, not that it cannot work. ChatGPT has more than 4,000 plugins, so check the plugin directory for a specific tool before you rule dots out.

Control and privacy compared

Five founder jobs and which agent fits each

This table is our reading of each product's documented features, plus our own experience with Grok Bot's output. Use it to decide which agent to trial for which job.

If you do only one thing with an agent this month, make it the support inbox. It is repetitive, it is easy to check, and good replies turn into revenue. Failed payments come second, because involuntary churn is often the easiest money a small SaaS can recover; our churn calculator shows what it costs you. New-user emails are a good third job once the first two work, and our guide to onboarding without a success team covers what those emails should say.

Set guardrails before you connect anything

Whichever agent you choose, set these up on day one:

1. Start read-only. Connect Stripe, your inbox and your analytics with the narrowest access available, and add write access only for jobs the agent has done well for a week.
2. Require approval for anything that leaves the business: sending email, publishing, spending, refunding, deleting and changing permissions.
3. Keep it away from production. No admin access to your app's database or hosting. Let it propose code changes as pull requests you review.
4. Keep secrets out of instructions. Assume any agent can repeat what you told it, as Grok Bot did in our use. Passwords, API keys and customer data never belong in a prompt.
5. Use a separate identity where you can. Muse has its own email address; for the others, consider a dedicated mailbox or account with limited permissions rather than your main admin login.
6. Treat every incoming email and web page as untrusted. Customers, spammers and competitors can all write text your agent will read, and approval rules are your defence against prompt injection.
7. Turn off model training where the plan lets you, especially if the agent reads customer data.
8. Review its work weekly. Look at what it sent, changed and spent, not just its summary of what it did.

If you have customers in the EU or UK, check that your privacy policy and data processing terms cover an AI agent reading their messages before you connect your support inbox. And before you automate anything, apply the rule from our guide to staying under 20 people: automate the work you repeat, not the work you have not figured out yet.

Test your pick in one week

Reviews, including this one, cannot tell you how an agent handles your business. A week of structured testing can, and each agent can be tried on a plan you may already have or can cancel monthly:

Which one to pick

The founders who get the most from these agents will not be the ones who hand over the most access. They will be the ones who give an agent one boring, checkable job, prove it works, and add the next job only after that. We will update this guide as prices, availability and limits change; all three companies were still adjusting them when we wrote it.