AI Secretary Comparison
Muse · dots · Grok Bot · Claude Duet · Open claw — Compare Muse, dots, Grok Bot, Claude Duet and the unresolved Open claw name by delegated work, autonomy, connections, price, memory and permissions. Different products, not equivalent capabilities.
A practical comparison of what you can hand off. “AI secretary” is a reader-friendly umbrella, not a claim that all five have identical features or availability. Open claw remains unidentified; OpenClaw is a candidate only. Reviewed Nov 4, 2026; no hands-on tests. LLM and decision-model hubs remain separate.
1. Muse
Verified Identified as Meta Muse through Meta’s announcement and product page. The agent product is distinct from foundation models such as Muse Spark.
| Field | Confirmed scope and limitations |
|---|---|
| Work you can delegate | Vendor claim Email, calendars, research, documents, purchases and goal tracking. Meta says work continues in the background after the app closes. |
| Autonomy and approval | Vendor claim Dedicated Muse Secure VM, system-separated Sentinel, approvals for actions such as messages and purchases, audit trail, per-connection permissions and revocation. Sentinel approval does not mean the user approves every action. |
| Connected tools | Verified Email, calendars, Instagram and other apps. Interfaces include the Muse app, WhatsApp and Mac. A complete connector availability matrix was not verified. |
| Pricing and availability | Verified Announcement: US rollout on iOS, Android and web. The current page also describes Mac, WhatsApp and Muse for Small Business; release dates for those additions are unconfirmed. Japan availability and age requirements were not verified. Official FAQ confirms a free usage allowance and paid subscriptions. Current paid amounts were not verified; the old article’s $20/$100 figures are not restated as current. |
| Memory and personal data | Verified Credentials are stored outside model visibility. Training opt-out, no sharing of conversations or VM data with Meta ad systems, and user-directed forgetting are documented. Confidential VM was announced for later in the year; shipment was not verified. |
| Permissions and safety | Vendor claim Dedicated Muse Secure VM, system-separated Sentinel, approvals for actions such as messages and purchases, audit trail, per-connection permissions and revocation. Sentinel approval does not mean the user approves every action. |
2. dots
Verified OpenAI uses lowercase dots, and dot for an individual agent. This is the product called Dots in the inherited article.
| Field | Confirmed scope and limitations |
|---|---|
| Work you can delegate | Vendor claim Dedicated cloud computer, ongoing projects and ChatGPT/Slack/Teams conversations. The Nov 3, 2026 update adds GPT-6.1 Terra to its model family; it does not claim a complete replacement of Astra. Specialist dots remain focused enterprise pilots. |
| Autonomy and approval | Vendor claim Custom Rules, Activity View and auto-review. Proactive research is restricted to read-only connected-app tools. The official page leaves password changes to the user, a narrower rule than early reporting suggested. |
| Connected tools | Verified ChatGPT, Slack and Teams conversations; connected apps and delegation to Codex/ChatGPT Work. Regional and plan-specific connector differences were not verified. |
| Pricing and availability | Verified Rolling out to Pro and Business Premium in eligible markets. Enterprise, including Edu and Healthcare, is an admin-enabled beta. The complete market list and Japan availability were not verified. The first dot is included in Pro or Business Premium at no extra cost, with an allowance for deeper work. Current plan subscription amounts were not verified. Dot conversations do not count toward ChatGPT limits; delegated Codex/ChatGPT Work tasks use their usual limits. |
| Memory and personal data | Verified Personal plans follow training settings. Business/Enterprise/Edu content is not used for training by default. Proactive research and self-notes are not trained on directly, but information incorporated into eligible conversations or tasks may be used depending on settings. |
| Permissions and safety | Vendor claim Custom Rules, Activity View and auto-review. Proactive research is restricted to read-only connected-app tools. The official page leaves password changes to the user, a narrower rule than early reporting suggested. |
3. Grok Bot
Verified Grok Bot is documented on x.ai; its pages also use the provider name SpaceXSI. Distinct from X reply bots and Grok foundation models. Corporate restructuring history was not investigated.
| Field | Confirmed scope and limitations |
|---|---|
| Work you can delegate | Vendor claim Persistent cloud computer, cooperating Bots, routines and event triggers. Oct 26 changes cover permissions, payment Vault and 1Password approvals; Oct 27 adds shopping cards, delivery tracking, Shop Pay and Team Bot sharing controls. |
| Autonomy and approval | Vendor claim Separate review agent, allow/block rules and escalation. The guide explicitly says users trust the model to follow rules. It also says other Bots can reach a site logged into by one Bot: do not assume complete per-Bot login isolation. |
| Connected tools | Verified Signed-in websites, MCP, plugins, skills and Cursor integration. Check configuration and documented boundaries for multiple accounts and Bot sharing. |
| Pricing and availability | Verified Launch announcement: beta for eligible SuperGrok and Cursor subscribers. Current product page lists inclusion with eligible Premium/Super/Ultra and other plans. Launch eligibility is not frozen as today’s terms. Region and age restrictions were not verified. Current monthly-billed USD prices: Premium $30/month, Super $100/month, Ultra $200/month, retrieved Nov 4, 2026. Weekly usage is included. Extra-use rates, taxes and regional differences were not verified. |
| Memory and personal data | Verified Conversation and workflow context can persist and be shared between Bots. Vault and 1Password credential controls are documented. Bot-specific training, retention and deletion boundaries were not verified. |
| Permissions and safety | Vendor claim Separate review agent, allow/block rules and escalation. The guide explicitly says users trust the model to follow rules. It also says other Bots can reach a site logged into by one Bot: do not assume complete per-Bot login isolation. |
Sources, retrieved Nov 4, 2026: bot ／ botintro ／ botguide ／ botlog
4. Claude Duet
Verified Anthropic release notes record the research preview on Oct 21, 2026. Claude Duet is an application; Opus 5.5 is its underlying model.
| Field | Confirmed scope and limitations |
|---|---|
| Work you can delegate | Vendor claim Morning briefs, inbox/calendar checks, drafts, invoice/payroll preparation, documents, decks and dashboards. Cloud work continues away from the device. Connected apps are checked every few hours, not guaranteed continuous real-time monitoring. |
| Autonomy and approval | Vendor claim Approvals before messages, purchases, signatures and deletion; separate action and sensitive-information classifiers. Securities trading and inter-account fund transfers remain off-limits even with approval. Credentials and browser are isolated from the model. |
| Connected tools | Verified Carries over connected apps and skills from the Claude account, including email, calendars and business services. Disconnection also applies across Claude. |
| Pricing and availability | Verified Current official page: waitlisted research preview for US Pro, Max and Ultra users; Team and Enterprise excluded. Standalone iOS/Android app and claude.ai. Ultra supplements the inherited article’s Pro/Max-only scope. Included in eligible plans during the preview, sharing Claude usage limits. Background checks count. Standalone pricing and current plan amounts were not verified. |
| Memory and personal data | Verified Carries over Claude memory, skills and connections. Disconnecting an app stops new access but learned context remains until memory reset. History deletion, reset and account deletion are documented. Duet-specific training policy and fixed retention days were not verified. Classifier misses and false positives remain possible. |
| Permissions and safety | Vendor claim Approvals before messages, purchases, signatures and deletion; separate action and sensitive-information classifiers. Securities trading and inter-account fund transfers remain off-limits even with approval. Credentials and browser are isolated from the model. |
Sources, retrieved Nov 4, 2026: duet ／ duetblog ／ duetsafe ／ duetlog
5. Open claw
Verified The supplied name “Open claw” could not be matched to a confirmed provider. OpenClaw at openclaw.ai is a candidate, not a confirmed identity. The following describes that candidate only; similarly named domains are not merged.
| Field | Confirmed scope and limitations |
|---|---|
| Work you can delegate | Vendor claim Self-operated agent for email, calendars, files and browser tasks through channels such as WhatsApp and Telegram. Neither a foundation LLM nor a bounded-answer decision model. |
| Autonomy and approval | Vendor claim Regular host Gateway defaults to loopback, unknown DMs use pairing, and groups use allowlists, with documented exceptions such as container exposed binds. Tool permissions, sandboxing and network exposure require configuration; local execution alone does not establish safety. |
| Connected tools | Verified Supplied identity unresolved. Candidate OpenClaw connects channels such as WhatsApp/Telegram, selected models and external services. Complete coverage and terms were not verified. |
| Pricing and availability | Verified Candidate official site describes open source, running on your machine, and independent nonprofit stewardship, with macOS/Windows/Linux apps. Its displayed v2026.10.10 is not asserted to be the latest release. Candidate site says no subscription and no hosted tier. This does not establish that model APIs, compute or messaging are free. License text was not checked in this review. |
| Memory and personal data | Verified Local models, hosted APIs and hybrid routing are supported. Self-hosting can still send data to model providers and connected services; it does not guarantee all data stays on-device. Retention and deletion depend on deployment and providers. |
| Permissions and safety | Vendor claim Regular host Gateway defaults to loopback, unknown DMs use pairing, and groups use allowlists, with documented exceptions such as container exposed binds. Tool permissions, sandboxing and network exposure require configuration; local execution alone does not establish safety. |
Before subscribing: four practical checks
Use these questions alongside the six comparison axes. The reference article by masa_wunder on note reports personal work experience with Muse, dots and Grok Bot; the author deferred trying Claude Duet. It is not a hands-on comparison of all five targets and does not identify Open claw.
Several roles, or one continuing counterpart?
The article uses agent count as a selection question. Using one Muse is the author's observation, not a verified product limit. Role instructions do not establish separate access boundaries. Reference article
Official check: Grok Bot supports collaborating Bots, but another Bot can reach a site logged into by one Bot. OpenAI describes a first dot and future additional dots. An official Muse agent-count limit was not verified. Grok Bot guide / OpenAI announcement
Can your actual tools and workflows run?
The article's preference for Grok Bot's flexibility is task-specific experience. Test your own apps, integrations, long-running jobs and recovery steps rather than treating it as a universal ranking. Reference article
Official check: Grok Bot describes running code and software; Muse describes building missing tools; dots describes app development and optional local-computer access. Execution is not exclusive to Grok Bot. Comparative CPU, memory, runtime limits and overall flexibility remain untested. Grok Bot / Muse / dots help
Does voice handoff work for you?
Voice preference, waiting time and response to interruptions are the article author's experiences, not a common-protocol benchmark. Test your device and language. Official dots help distinguishes user-initiated calls from the inability of a dot to initiate calls at launch. Reference article / Official dots help
Do allowances and regional terms fit, beyond the entry price?
The article offers a shortlist, not proof of a universally cheapest product or lower cost than hiring. Personal use from Japan and reported Japanese integrations do not guarantee availability to every account. Reference article
Official check: Muse has a limited free allowance. The first dot is included in Pro or Business Premium at no extra cost, subject to rollout and eligibility. Grok Bot displays monthly-billed USD Premium $30, Super $100 and Ultra $200, checked Nov 4, 2026. Claude Duet is a waitlisted US research preview for Pro, Max and Ultra, sharing Claude limits. Check taxes, regional differences, additional usage and required integrations separately. Muse / dots / Grok Bot pricing / Claude Duet
Evidence boundary: Selection questions are summarized, without reproducing the article's prose or illustrations. Important capabilities, prices and product identities were checked against vendor sources, separately from the author's opinions. Direct URLs for the official screenshots mentioned in the article were not visible in the retrieved text; none were guessed. The article's formal title and publication date were not verified. Accessed Nov 4, 2026. Reference article
6. Gemini Spark (Google) — added 2026-11-06
An always-on personal AI agent announced by Google on 19 May 2026. It is a feature of the Gemini app; no standalone product called a "secretary model" was confirmed.
- Runs on Gemini 3.5 with the Google Antigravity harness, in the cloud, so work continues with the device closed.
- Workspace integration (Gmail, Docs, Slides); MCP integrations with Canva, OpenTable and Instacart announced.
- User-controlled activation and connections; confirmation required before payments or sending mail.
- Availability at launch: trusted testers, then a US Google AI Ultra beta; Japan rollout to Pro users from 29 July 2026 (third-party report).
Source: Google Japan Blog, 19 May 2026 https://blog.google/intl/ja-jp/company-news/technology/next-evolution-gemini-app/ (retrieved 2026-11-06)
- The six-subject structure — Muse / dots / Grok Bot / Claude Duet / Open claw / Gemini Spark — is retained. No new officially confirmed change to availability, pricing, features or regions was found.
- Unverified Gemini Spark's Japan availability, standalone pricing and enterprise controls remain unconfirmed on Google's official pages; the Japan rollout appears only in third-party reporting (July 2026).
- Unverified The identification caveat for Open claw (OpenClaw as candidate) is retained.
- Prices and performance figures are carried over; nothing unverified was restated as a new price.
Choosing by operational boundary
AIfrontier viewCompare actions, not model scores. Before connecting accounts, check read/write separation, approval rules, shared credentials, budget, logs, and deletion of learned context after disconnect. A cloud computer, a reviewer and self-hosting are different boundaries; none alone establishes safety.
Third-party evidenceWIRED’s launch report documents dots and the personal-agent trend, but its launch-era price and approval examples are not used as current official terms. WIRED
Sources and update log
- meta — https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
- muse — https://ai.meta.com/muse/
- dots — https://openai.com/index/introducing-dots/
- bot — https://x.ai/bot
- botintro — https://x.ai/news/introducing-grok-bot
- botguide — https://x.ai/bot/guides/grok-bot-101
- botlog — https://x.ai/changelog/bot
- duet — https://claude.com/product/claude-duet
- duetblog — https://claude.com/resources/articles/claude-duet
- duetsafe — https://claude.com/resources/articles/claude-duet-safety-privacy-and-security
- duetlog — https://docs.anthropic.com/en/release-notes/claude-apps
- claw — https://openclaw.ai/
- clawsafe — https://docs.openclaw.ai/gateway/security
- clawlocal — https://docs.openclaw.ai/gateway/local-models
- wired — https://www.wired.com/story/openai-dots-always-on-ai-agents-that-proactively-help/
2026.11.06: Created this five-subject standing hub inside Web v94. Existing Oct 3 and Oct 22 articles remain historical; targeted official updates and links were added. No blanket re-verification of their earlier content is claimed.
2026.11.06: Renamed the existing hub AI Secretary Comparison and reorganized the same evidence under six practical axes. File paths, source dates and unresolved identity notes are unchanged.