# Always-On AI Agents Compared for Gemini Notebook (formerly NotebookLM)
> Dots vs Muse vs Grok Bot vs Claude Duet vs Gemini Spark for research work: prices, approvals, privacy, and which of them can read your NotebookLM library.
By Juli · Published Oct 26, 2026 · https://www.kortex-notebooklm.com/blog/always-on-ai-agents-compared
Five companies sell the same promise: an AI that keeps working after you close the laptop. Always-on AI agents like Claude Duet, ChatGPT Dots, Gemini Spark, Meta Muse and Grok Bot each get their own cloud computer, check your accounts in the background and message you first. For research and knowledge work, though, the questions are narrower. Can it read your sources, cite them, brief you from your documents and keep a project's context? What does it do without asking? And can it open the library you've built in Gemini Notebook (formerly NotebookLM)? This guide compares all five on exactly that.
The pull is real. When we analyzed the 776 answers on the Kortex MCP waitlist survey (Sep-Oct 2026, then in private beta), the most requested job was running notebook work on a schedule or trigger, asked for by 128 people. One respondent put the always-on idea in a single line: "Each morning, my assistant should prepare a short source-backed briefing from selected notebooks". Search shows people weighing the options. When we pulled Google autocomplete on October 26, 2026, "dots vs muse" completed to "which is better" and "vs grok bot", "claude duet" completed to "waitlist", and "chatgpt dots" completed to "price" and "plus plan".
> **Key Takeaways**
> - All five agents run on their own cloud computer and ask before consequential actions, but they differ on plans, regions and what they'll never do.
> - For research, Claude Duet and ChatGPT Dots lean toward work, Meta Muse toward personal errands, Grok Bot toward teams of role-based bots.
> - Gemini Spark (called Task in the Gemini app) can add a Gemini Notebook to a task natively ([Gemini Apps Help](https://support.google.com/gemini/answer/17094507?hl=en), 2026).
> - Claude and ChatGPT reach Gemini Notebook through Kortex MCP, included with Kortex Pro and Lifetime. Muse and Grok Bot have no documented route.
## What is an always-on AI agent?
An **always-on AI agent** is an assistant with its own cloud computer that keeps working in the background, checks your connected apps on its own, and comes to you when something needs a decision. In Anthropic's words, Claude in Duet "comes to you with what needs your attention, completes tasks you ask of it, keeps working even when you step away, and checks with you before it acts on your behalf" ([Anthropic](https://claude.com/resources/articles/claude-duet), 2026). OpenAI calls a dot "an always-on agent that keeps work moving and asks for your input when a decision needs you" ([OpenAI dots docs](https://learn.chatgpt.com/docs/dots), 2026).
That's the shared shape. The differences sit underneath: who each one is built for, how it reads your accounts, and where it stops to ask. One naming note before the table. Google's help center calls Gemini Spark **Task** inside the Gemini app; Google says it chose the word because users kept asking for help finishing tasks (Gemini Apps Help, 2026). People still search for Gemini Spark, so this guide uses both names.
The closest relative most readers already know is a chat assistant with connectors. The difference is initiative. A chat waits for you. An always-on agent wakes itself up, reads what changed, and decides whether you need to hear about it.
## Dots vs Muse vs Grok Bot vs Duet vs Spark at a glance
Here are the basics from each maker's own pages. Prices are US list prices for the cheapest plan that includes the agent.
| Agent | Maker and model | Who gets it | Cheapest way in (US) |
|---|---|---|---|
| Claude Duet | Anthropic, Claude Opus 5.5 | Research preview by waitlist, Pro and Max in the US | Claude Pro, $20 a month ($17 billed yearly) |
| ChatGPT Dots | OpenAI, GPT-6 Astra | Pro (18+, outside the EEA, UK and Switzerland), Business Premium, Enterprise | ChatGPT Pro, from $100 a month |
| Gemini Spark (Task) | Google, Gemini models | Personal accounts, 18+, not in the EEA, Nigeria, Switzerland or the UK | Google AI Pro, $19.99 a month |
| Meta Muse | Meta, Muse Spark | US and Canada, in the Muse app, on muse.ai and in WhatsApp | Free with a usage limit; Power is $20 a month |
| Grok Bot | xAI, model not named in its docs | Desktop and mobile apps, through Cursor or SuperGrok plans | Cursor Premium, $30 a month |
Sources: Claude Help Center, [Claude pricing](https://claude.com/pricing), OpenAI dots docs, Gemini Apps Help, [Meta](https://about.fb.com/news/2026/09/introducing-muse-small-business/), [Meta subscriptions](https://www.meta.com/help/subscriptions/1021145227643680/), [xAI](https://docs.x.ai/grok-bot/faq), [Cursor pricing](https://cursor.com/pricing), 2026.
Two of those rows need a footnote. Duet is included in Pro and Max at no extra charge, but it "uses your plan's usage limits, including the five-hour session and weekly limits", and its background work like the morning brief counts too (Anthropic, 2026). OpenAI goes the other way: "Conversations with your dot don't count toward your ChatGPT usage limits," although tasks it starts in Work or Codex do (OpenAI, 2026).
## Which always-on agent is best for research and knowledge work?
None of the five is a research tool first. Each is a general agent that can do research when you point it at sources. So the useful question is what each one reads, what it returns, and what it remembers between jobs. Here's how they compare on that, one at a time.
### Claude Duet: briefs and files for professionals
Anthropic built Duet for "professionals and small business owners with more on their plates than they can track" (Anthropic, 2026). Each morning it looks across your email and calendar for priorities. Through the day it checks your accounts "at set intervals, so something urgent can arrive between checks". You talk to it in one ongoing thread, or open channels for separate topics that "share the same memory". For research, the useful part is output. "Bigger results come back as files you can review and edit, such as a research brief", and finished work lands in the **Make** tab ([Claude Help Center](https://support.claude.com/en/articles/17631495-get-started-with-claude-duet), 2026).
Context carries over from Claude: "the memory, skills, and connectors you've set up in Claude are there from your first message". There's one catch for project work, though. Duet "can't see all of your projects or the files on your computer, and it can read your work in Claude but can't change it" (Claude Help Center, 2026). So a carefully built Claude Project doesn't automatically become Duet's library. Every's early review, written with access Anthropic provided, called Duet a stronger professional tool than OpenAI's dot, but noted that "its scheduled runs can't reach my custom connectors" ([Every](https://every.to/vibe-check/vibe-check-duet-anthropic-s-always-on-agent), 2026).
### ChatGPT Dots: the orchestrator inside ChatGPT
A dot lives inside ChatGPT and reaches you there, in Slack, in Microsoft Teams or by phone call. Its research mode is cautious by design: it researches connected information, and "That research is read-only; follow-up actions require the applicable permissions" (OpenAI dots docs, 2026). It can run background agents in parallel and start Work or Codex tasks, which makes it strongest when your projects already live in ChatGPT and you want one agent keeping them moving.
Memory is where research users should look closely. A dot keeps "its own saved notes about your preferences, decisions, and ongoing work" (OpenAI dots docs, 2026). OpenAI's privacy FAQ says you can't view, delete or directly edit individual dot memories, and the way to clear them is to delete the dot. Disconnecting a plugin stops new access but doesn't remove what the dot already learned ([OpenAI Help Center](https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs), 2026). For a project with sensitive sources, that's worth knowing before you connect them.
### Gemini Spark (Task): the one closest to your notebooks
Google's help page says you can "Assign Gemini tasks to automate complex workflows and manage schedules for ongoing tasks", with up to 15 tasks running at once (Gemini Apps Help, 2026). It uses Google services directly, and Gemini Notebook is on the list next to Drive, Docs, Gmail, Keep and YouTube. To hand it a notebook, choose **Upload & tools**, then **Add Notebook**. It runs in the Gemini mobile app, the Gemini app on Mac and gemini.google.com, and it needs Google AI Pro or Ultra, a personal account, Keep Activity on and Personal Intelligence opted in.
That native link is the strongest research advantage of the five. One caveat carries over from the Gemini app, however. Google says Gemini Notebook answers are grounded only in your sources, while in the Gemini app they "may also include web search and other tools", as our guide to [Gemini Notebook in the Gemini app](/blog/gemini-notebook-in-gemini-app) explains. The gap is measurable. In a 2025 study on a 300-document corpus, NotebookLM hallucinated in 13% of outputs, against 40% for Gemini and ChatGPT ([Hagar, Agustianto & Diakopoulos](https://arxiv.org/abs/2509.25498), 2025). Both apps have changed since, but the design difference hasn't. For a quote you'll cite, check it in Gemini Notebook itself.
### Meta Muse: the free personal assistant
Muse is the only one of the five with a free tier: it's free with a usage limit. Meta built it for everyday life: it runs on a dedicated cloud machine Meta calls the Muse Secure VM, remembers what matters to you, and you can tell it to forget specific things ([Meta](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/), 2026). It reads and creates **artifacts** such as "documents, images, code and web apps", saved to your Library ([Muse Help Center](https://www.meta.com/help/superintelligence/2074655449783957/), 2026).
For knowledge work, Meta added connectors aimed at small businesses, including Notion, Dropbox, Box, Granola, Slack and Zoom (Meta, 2026). Many connectors can be set to read only, so Muse can "retrieve data, but not take actions" ([Muse Help Center](https://www.meta.com/help/superintelligence/1687253048996149/), 2026). It's a capable research helper for personal topics. It isn't built around citing a library of documents, and Meta's pages don't describe citation features.
### Grok Bot: a team of bots with jobs
Grok Bot gives you several Bots with names, jobs and memory, which can message each other and hand work off ([xAI](https://docs.x.ai/grok-bot/overview), 2026). For research, that means one Bot can own a recurring digest while another owns sourcing. xAI's docs are unusually concrete about evidence: they suggest asking for "A document with headings and source links" and having the Bot separate "Facts found in source systems" from assumptions ([xAI](https://docs.x.ai/grok-bot/files-and-results), 2026). Bots can also build **Pages**, styled reports and dashboards.
Two cautions come straight from those docs. On memory: "For important decisions, ask the Bot to check the current source rather than relying on memory". On separation, all your Bots share one cloud computer, files and browser logins included. So "Do not use separate Bots as a security boundary" ([xAI](https://docs.x.ai/grok-bot/approvals-security-and-privacy), 2026).
### The verdict for research
If your sources already live in Gemini Notebook, Gemini Spark is the only agent that opens them natively. If your work lives in Claude or ChatGPT, Duet and Dots are the stronger agents, and Kortex MCP gives either one your notebooks. Grok Bot fits recurring, role-based research where you want source links on every claim. Muse is the one you can try for free, but it's built as a personal assistant, not a research library.
Whichever you pick, it finds notebooks the way a new colleague would: by their names. Gemini matches your words to notebook titles, so three notebooks called "Research" and one "Untitled notebook" become a guessing game. That's common among people who lean on notebooks. In the Kortex MCP waitlist survey, 327 of 776 respondents (42%) kept more than 50 notebooks.
Name and tag your notebooks before an agent goes looking
Kortex shows your whole Gemini Notebook library in one table, with source counts, tags and collections, so vague names and duplicates stand out before any agent has to guess. Free on Chrome.
## What does each agent do without asking?
Reading and drafting, mostly. All five act on their own for low-risk work and stop for anything that sends, spends or deletes. The real differences are in the hard limits, and in how easy it is to turn a one-off approval into a standing one.
| Agent | Stops to ask before | Won't do | Memory and data controls |
|---|---|---|---|
| Claude Duet | Sending a message, making a payment, signing a document, deleting anything | Trade securities, move money between accounts, access someone else's account, call or text people | It can search past Duet chats and non-project Claude chats; turn off "Search and reference chats" |
| ChatGPT Dots | Whatever OpenAI's automatic review flags, plus your own instructions and Custom Rules | Let your instructions or Custom Rules override built-in safety requirements | You can't view or edit single memories; deleting the dot clears them |
| Gemini Spark (Task) | Sending communications, changing or deleting data, purchases, financial transactions, web forms | Share health, financial or ID details without pausing, even on "Always allow" | Delete saved remote-browser data in Settings, under Browser |
| Meta Muse | Sensitive actions like payments or purchases; every write action on connectors set to Ask | Not specified in Meta's help pages | Model training on your chats is on by default; turn it off in Data controls |
| Grok Bot | Actions flagged by Auto Review and your own "Ask first" rules | Type your passwords; it hands those steps to you | Training opt-out follows your Cursor account settings |
Sources: Claude Help Center, OpenAI dots docs and Help Center, Gemini Apps Help, [Muse Help Center](https://www.meta.com/help/superintelligence/1047255454427887/), xAI docs, 2026.
The setting to watch is "always allow". Muse's approval card offers **Allow once** and **Always allow**, and the second lets Muse "take this type of action for this Connector in the future without asking again" ([Muse Help Center](https://www.meta.com/help/superintelligence/1385290430137537/), 2026). Grok Bot's **Always allow** can save a matching rule too. That's convenient for reading a folder. It's risky for anything that sends a message or shares an address. A good habit across all five: approve once while you learn what the agent does, and save standing rules only for read-only actions.
Every one of these agents also reads text other people wrote. **Prompt injection** is text hidden in an email, web page or document that tries to give the agent instructions of its own. Anthropic says Duet "is built to treat what it reads as information and to act on what you ask", and its help page admits no defense is perfect (Claude Help Center, 2026). Keep that in mind before connecting an inbox and a research library to the same agent.
## How can each agent reach your Gemini Notebook library?
Gemini Notebook has no public consumer API, so an agent can't simply log in and ask your notebooks. That leaves three routes: Google's own connected app, which only Gemini has; an MCP connector such as Kortex MCP, where the agent accepts one; or a file you hand over. **MCP** (Model Context Protocol) is the open standard AI apps use to call outside tools, and our [Gemini Notebook MCP guide](/blog/notebooklm-mcp) explains how it works.
| Agent | Route to Gemini Notebook | Status |
|---|---|---|
| Gemini Spark (Task) | Add Notebook to a task, or type @Gemini Notebook in a chat | Native, from Google |
| Claude Duet | Kortex MCP, added in Claude as a custom connector | Carries over with your Claude connectors; Anthropic doesn't single out custom ones |
| ChatGPT Dots | Kortex MCP, added in ChatGPT as an MCP app | Works in ChatGPT; OpenAI's dots docs say "supported plugins" without listing custom ones |
| Meta Muse | None documented | Custom connectors are built against a public API, which Gemini Notebook doesn't offer |
| Grok Bot | None documented for your own MCP server | Catalog apps only in xAI's docs; export files instead |
The routes line up with what people asked for. In the waitlist survey, respondents named Claude 126 times, ChatGPT 69 times and Gemini 61 times. Grok came up 14 times, and Meta's assistant wasn't among the ten most named.
Assistants named, out of 776 respondents (one person could name several). In bold: makers of agents compared here. Source: Kortex MCP waitlist survey, Sep 19 to Oct 1, 2026.
### Gemini Spark: add the notebook to the task
This route is built in. In task mode, choose **Upload & tools**, then **Add Notebook**, and Gemini works with that notebook as a source for the task (Gemini Apps Help, 2026). In an ordinary chat, type **@Gemini Notebook** and name the notebook. Give the task a job and a notebook by name, like this:
```
Every Monday morning, read my "Q4 market research" notebook, list what the sources say changed since my last brief, and tell me which sources each point came from.
```
Gemini also accepts custom connected apps by MCP server URL, for personal accounts in the US, in English ([Gemini Apps Help](https://support.google.com/gemini/answer/17209137), 2026). Kortex's client guides don't cover Gemini, though, and the native route already does the core job.
### Claude Duet and ChatGPT Dots: Kortex MCP as a connector
Claude and ChatGPT have no Google door, so they need a bridge. **Kortex MCP** is that bridge: your agent calls one URL, and each call runs through the Kortex extension in your own Chrome, which asks Gemini Notebook with the Google session you're already signed into ([Kortex MCP docs](/docs/mcp-server), 2026). The answer comes back with Gemini Notebook's own citations, and `ask_library` can ask up to three notebooks at once, each answer kept separate.
Setup takes a URL and a sign-in. Our guide to [connecting Gemini Notebook to Claude and ChatGPT](/blog/gemini-notebook-in-claude) has every step.
1. **Claude:** go to **Customize > Connectors**, choose **Add custom connector**, paste `https://mcp.kortex-notebooklm.com/mcp` and sign in on the Kortex consent page. Anthropic offers custom connectors on Free, Pro, Max, Team and Enterprise, with one on Free ([Claude Help Center](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp), 2026).
2. **ChatGPT:** add Kortex as an MCP app with the same URL and OAuth sign-in, through developer mode ([OpenAI](https://developers.openai.com/api/docs/guides/developer-mode), 2026).
3. **Then check from the agent itself.** Neither maker documents custom connectors inside its agent specifically, so ask Duet or your dot directly:
```
List the Kortex tools you can use. Then ask my "Client research" notebook through Kortex what the three main risks are, quote its citations, and label anything that comes from you rather than the notebook.
```
If the tools show up, the agent can work from your notebooks like any other app. If they don't, the same prompt works in an ordinary Claude or ChatGPT chat, and in Claude Cowork.
Give Claude and ChatGPT your notebooks as their library
One connector, and Claude or ChatGPT can ask your Gemini notebooks and bring back answers with their citations attached. Kortex MCP is included with Kortex Pro and Lifetime.
## What changes when the agent works while you sleep?
The agent keeps running in the cloud; your notebook bridge doesn't. Kortex MCP relays every call through the Kortex extension, so Chrome has to be open and signed in, and the first call after an idle spell can take up to 30 seconds (Kortex MCP docs, 2026). A 6 a.m. brief that asks your notebooks while the laptop is shut gets nothing back from them.
Plan around it instead of fighting it. Three habits keep an always-on agent and a browser-based bridge working together:
- **Time notebook jobs for your working hours.** Ask the agent to brief you from your notebooks at 9 a.m., not 6.
- **Let Kortex do the scheduled notebook part.** Kortex schedules run hourly, daily or weekly from Chrome, and can ask a notebook and save the answer as a note while your computer is on. Our guide to [scheduled Gemini Notebook briefs](/blog/gemini-notebook-daily-briefing) shows the recipes.
- **Keep risky tools on ask.** In Claude's connector settings, each tool can be set to always allow, ask every time or never. Leave Kortex's reading tools on allow and anything that deletes on ask.
What about Muse and Grok Bot? Without a connector, hand them a file. Grok Bot takes source material you attach, Muse can read and summarize documents, and Kortex exports a notebook's sources, notes and chats as PDF, Markdown, text or ZIP. You lose live citations and follow-up questions that way, but the agent gets a fixed copy of your research to work from.
In short: Gemini opens your notebooks itself, Claude and ChatGPT open them through Kortex MCP, and the other two get a file.
Let your agent ask your notebooks, citations intact
Kortex MCP connects Claude, ChatGPT and other MCP clients to every notebook in your Gemini Notebook library, with Gemini Notebook's citations passed back as returned. It's included with Kortex Pro and Lifetime.
## Frequently asked questions
### Claude Duet vs Grok Bot: which is better for research?
It depends on how you work. Claude Duet is one agent built for professionals: a morning brief, topic channels and research briefs returned as files. Grok Bot is a team of Bots with separate jobs that share one cloud computer, and xAI's docs suggest asking for documents with source links. Only Duet can reach Gemini Notebook (formerly NotebookLM), through Kortex MCP as a Claude connector.
### ChatGPT Dots vs Meta Muse: what's the difference?
A dot lives inside ChatGPT, runs on GPT-6 Astra and coordinates work across ChatGPT, Work and Codex. It needs ChatGPT Pro, Business Premium or Enterprise. Muse is Meta's personal agent in its own app, on the web and in WhatsApp, free with a usage limit in the US and Canada. Dots suit work projects; Muse suits personal errands.
### Gemini Spark vs Claude Cowork: which should I use?
Gemini Spark, which Google calls Task in the Gemini app, automates multi-step work, runs ongoing tasks on a schedule and can add a Gemini Notebook to a task natively. Claude Cowork is the agent mode inside the regular Claude app, and on Pro and Max its sessions run in the cloud. Cowork reaches your notebooks through Kortex MCP as a custom connector.
### Is there a Claude Duet waitlist?
Yes. Claude Duet is a research preview for Claude Pro and Max subscribers in the US, and you join a waitlist to get in. Anthropic emails you when it's ready for your account. It's included in Pro ($20 a month, or $17 billed yearly) and Max (from $100 a month) and uses the same usage limits as the rest of Claude.
### How much does ChatGPT Dots cost, and is it on the Plus plan?
Dots come with ChatGPT Pro, which OpenAI sells at three levels starting at $100 a month, plus Business Premium and Enterprise (where an admin turns them on). Plus isn't on OpenAI's list. On Pro, dots are for users over 18 outside the EEA, the UK and Switzerland, and conversations with your dot don't count toward ChatGPT usage limits.
### How much do Gemini Spark, Meta Muse and Grok Bot cost?
Gemini Spark (Task) needs Google AI Pro, $19.99 a month in the US, or Ultra. Meta Muse is free with a usage limit, or $20 a month for Power and $100 for Maximum. Grok Bot comes with Cursor's paid plans and linked SuperGrok subscriptions; Cursor's pricing page lists it from the $30-a-month Premium plan.
### Can always-on AI agents read my NotebookLM notebooks?
Some can. Gemini Spark reaches Gemini Notebook (formerly NotebookLM) natively: add a notebook to a task, or type @Gemini Notebook in a chat. Claude and ChatGPT reach it through Kortex MCP, included with Kortex Pro and Lifetime. Meta Muse and Grok Bot have no documented route, so give them exported files instead.
---
*Editorial note: fact-checked against the official help pages and docs linked above on October 26, 2026. Spotted something out of date? [Contact us](/support) and we will fix it, or read more [about Kortex](/about).*