Guide / Choosing your tools
Compare personal AI assistants
Choose by practical fit: Dots, Claude, Gemini, Microsoft Copilot, OpenClaw, and Meta Muse, with dated sources and clear limits.
Reviewed October 14, 2026. Evidence: vendor documentation, not a hands-on comparison. Just the Dots is an independent Blisster Inc publication. We have not run equivalent tasks across these products, measured reliability, or verified every account’s access. A passing website test says nothing about an assistant’s performance.
Why a dot might fit
If you already use ChatGPT to think, Codex to build, and Work to carry out tasks, a dot offers a way to coordinate an ongoing responsibility across that work. OpenAI documents using available ChatGPT context, following up on existing Codex threads, and reading or editing Work automations. That continuity is a practical reason to consider it—not evidence of better results for every job. October update.
You may be better served by another assistant’s app integrations, memory controls, or deployment options. Compare the particular product and account below. Consumer policies are not interchangeable with workplace or API contracts. “Not established here” means we lack evidence for that claim, not that the feature cannot exist.
Start with the work you already do
These are our editorial fit suggestions, based on the documentation linked below. They are not performance rankings.
| Assistant | Consider it when… | A trade-off to check |
|---|---|---|
| OpenAI Dots | Your ongoing work already spans ChatGPT, Codex, and Work. | Selective recall; no individual dot-memory inspection or editing; limited rollout. |
| Claude / Cowork | You want research and editable deliverables, with memory topics you can inspect and change. | Cloud and local tasks differ; desktop access needs the app open; multi-step work uses more allowance. |
| Google Gemini | Your useful context is in Google apps and you want it available for personal tasks. | Task mode requires Keep Activity and Personal Intelligence; connected-data training needs careful review. |
| Microsoft Copilot | You work with Office documents, Outlook, and OneDrive and want recurring assistance there. | Personal Cowork is a preview; local-file access and some file operations are unavailable. |
| OpenClaw | You want to choose hosting, models, tools, and channels—and can maintain the system. | You operate the infrastructure; self-hosting does not make connected services or model calls private. |
| Meta Muse | You want ongoing personal errands, connected-app help, and a persistent cloud browser. | Training starts enabled; some Meta accounts connect automatically; deletion and forgetting differ. |
Compare the controls, not just the promise
Open an assistant below. Each uses the same six criteria. Product claims are vendor-documented, not hands-on findings. Account, region, rollout, version, and administrator settings can change what you see.
OpenAI DotsPersonal ChatGPT plans; workplace differences noted
- Capabilities and practical fit
Coordinate ongoing responsibilities, delegate tasks, and return with changed priorities. Schedules, supported event monitoring, and contextual suggestions are different arrangements. This can fit work already underway in ChatGPT, Codex, and Work; it does not mean every thread or source is accessible, or that follow-up will always be useful.
- Availability and access permissions
Gradual rollout: adult Pro 100/200/500 outside the EEA, UK, and Switzerland; Business Premium worldwide; Enterprise requires administrator enablement. Mobile app and desktop setup are supported, not mobile web. Plugins and personal-computer access are separate permissions. Local work needs the connected computer online with ChatGPT open; cloud work can continue independently. Work/Codex allowances still apply.
- Action approvals
Automatic review considers your authorization, app permissions, optional Custom Rules, and built-in safeguards. Some steps proceed, some require approval, and some must be handed to you. A scoped authorization can cover later actions; drafting permission is not sending permission. Custom Rules do not grant app access and are not infallible.
- Visibility and stopping
Activity exposes delegated tasks and results; Scheduled shows recurring work, and the cloud browser supports takeover. Pause stops the main task, not all delegated work or future runs. Stop delegated tasks in Activity and disable/delete recurring work in Scheduled separately. Completed actions are not undone.
- Training policy
For personal plans, Improve the model for everyone governs eligible dot conversations and work, including connected information used there. Proactive research and private notes are not trained on directly; information carried into eligible work follows its data settings. Business/Enterprise/Edu data is not trained on by default. Limited human review can still occur when model improvement is off.
- Memory and deletion controls
Dot notes and ChatGPT memory are separate. OpenAI currently says you cannot inspect, delete, or directly edit individual dot memories. Turning off ChatGPT Memory stops sharing; disconnecting a plugin stops new access. Neither erases existing dot context. Deleting the dot deletes its context, but its separately stored files, Codex threads, ChatGPT conversations, and ChatGPT memories need separate management.
Claude / CoworkPersonal Pro/Max; unified experience is rolling out
- Capabilities and practical fit
Cowork combines research, connected tools, and file work to produce documents, spreadsheets, and presentations. Cloud tasks share chat memory and can continue across devices. This is a practical alternative for turning scattered material into deliverables. Multi-step tasks can consume substantially more usage than a short chat.
- Availability and access permissions
Personal Cowork requires Pro or Max. New cloud tasks span desktop, web, and mobile; older tasks may remain local. Desktop must stay open for local file/browser/computer access. Computer use is a Pro/Max beta with per-app permission and an app blocklist; screen interaction is not sandboxed from those applications. Connected folders and tools determine other access.
Anthropic: cloud and local workAnthropic: computer useAnthropic: Cowork guide
- Action approvals
Unified Claude offers Manual by default and Auto with automated action checks. Connector permissions still matter: Always allow is different from Needs approval or Blocked. Older Cowork interfaces can also expose Skip, which removes approval checks. Do not assume the same defaults across these experiences. Automated review is not a guarantee against mistakes.
- Visibility and stopping
You can review progress and outputs, stop work, or redirect it. Scheduled has past/upcoming runs and separate pause, edit, and delete controls. New cloud schedules can run while the device is off; work requiring local access still depends on the relevant computer. We have not verified an exact stop-button procedure or every internal action’s visibility.
- Training policy
Personal-plan model improvement is a user choice, with safety-review and explicit-feedback/consent exceptions. Opted-in data can be retained in training pipelines for up to five years. Opting out or deleting a chat affects future training; it does not reverse training already completed or underway. These are personal-product terms, not Anthropic’s commercial contract.
- Memory and deletion controls
Cloud Cowork shares Claude memory. You can inspect, edit, or delete memory topics, pause memory, or reset it. Deleting a conversation does not delete generated memories. Ordinary deleted chats leave visible history immediately and backend storage within 30 days, subject to policy exceptions. Topic-level control is a documented strength; it is not proof of better privacy overall.
Google GeminiPersonal account · Task mode, formerly Spark, and Personal Intelligence
- Capabilities and practical fit
Task mode can combine research, documents, inbox/calendar work, and website actions. Personal Intelligence brings relevant connected Google information into the task. Schedules support time, Gmail-filter, and topic triggers. This is a practical fit for Google-centered work; approximate timing makes schedules unsuitable for urgent detection.
- Availability and access permissions
Task mode requires an adult personal account, AI Pro/Ultra, Keep Activity on, and Personal Intelligence opt-in; listed regional exclusions include the EEA, UK, Switzerland, and Nigeria. Connected Apps and browser access are distinct. Local Chrome can include signed-in sites. Ordinary Gemini availability is broader than task-mode eligibility; check your account before planning a workflow.
- Action approvals
Google describes confirmation for communications, modifications, purchases, forms, and browser tasks, with takeover when needed. A connected source does not itself establish a schedule or event trigger. These safeguards are not infallible.
- Visibility and stopping
The work panel shows steps and files, with browser takeover. Stopping Chrome browsing can leave the task using another tool: use Stop in Gemini’s task text box to cancel the task. Manage schedules separately. Turning task mode off pauses tasks/schedules; turning it back on resumes them. We do not promise every task runs with your device off.
- Training policy
With Keep Activity on, Google permits improvement/training uses of prompts, responses, and connected-data summaries, excerpts, and inferences; these can sometimes include a whole relevant short email or file. Google distinguishes that from directly training on an entire inbox or library. Human review can occur. Task mode requires Keep Activity, so check this requirement before connecting sensitive sources.
- Memory and deletion controls
Google’s guidance for remembered information includes deleting chats containing it and disconnecting its source app. Deleting task activity removes associated schedules, but not generated files or remote browser/computer data. Manage those separately in Task settings. Turning task mode off deletes remote data but pauses rather than deletes tasks/schedules; chat activity has its own controls.
Microsoft CopilotUpdated personal app · Cowork Preview, replacing personal Tasks
- Capabilities and practical fit
Cowork documents Office-file creation/editing, email/calendar work, OneDrive, research, and scheduled briefings. Personal Tasks are migrating to Cowork. Its Microsoft workflow coverage is a practical reason to consider it; personal-account Preview and work/school Cowork are different scopes, not interchangeable feature or policy promises.
- Availability and access permissions
An eligible Microsoft 365 subscription is required for personal Cowork Preview. Authorized connectors supply account access; disconnecting revokes it, while toggling one off only pauses its use. Current documentation excludes direct access to local-device files and deletion of OneDrive/SharePoint files or folders. We have not established a complete personal-plan or region matrix here.
Microsoft: personal Cowork PreviewMicrosoft: Cowork data and access
- Action approvals
Action previews and explicit approval or takeover cover sensitive actions such as payments, communications, external personal information, and account changes. Some recurring actions can be approved during setup. Review the specific authorization rather than assuming you will receive a fresh prompt on every scheduled run.
- Visibility and stopping
The side panel shows progress and input/output files. You can pause immediately or after the current step, resume, or cancel. Confirmed schedules have separate edit, pause, and cancel controls in My tasks or the side panel. This makes review possible; it does not prove a task succeeded or reverse completed actions.
- Training policy
For the updated app with a personal Microsoft account, Microsoft says prompts, responses, file contents, and optional feedback are not used to train foundation models. Cowork screenshots are also excluded from training. Limited retention and some human/automated review still apply. This is a specific documented policy, not a claim of zero data use or a guarantee of privacy superiority.
Microsoft: updated-app activity policyMicrosoft: Cowork data and access
- Memory and deletion controls
Saved memories and chat history have separate deletion controls. Turning memory off does not delete existing saved memories. Use the current privacy controls and inspect the saved state after deleting. These updated-app policies must not be confused with older Copilot training defaults or retention figures; an exact Cowork retention period was not established in this review.
Microsoft: current privacy controlsMicrosoft: Cowork data and access
OpenClawSelf-hosted software · behavior depends on version, runtime, and configuration
- Capabilities and practical fit
An open-source gateway connects messaging channels with agents, models, browsers, memory, and automation. Model and deployment choice are practical strengths. It is not one uniform hosted assistant: the operator manages updates, credentials, configuration, and the running host, and installed tools determine what it can do.
- Availability and access permissions
Access follows channel authorization, tool policy, host permissions, sandboxing, and runtime. The gateway’s trust model assumes one trusted operator or mutually trusting users, not hostile tenants. Restrict message tools explicitly when cross-conversation or cross-provider communication is unwanted. Account subscriptions and third-party permissions remain separate from installing the software.
- Action approvals
Execution approvals are configurable. Full-permission paths can bypass host approval checks, so there is no universal promise that an OpenClaw agent will ask before acting. Review the actual installation’s approval modes, allowlists, fallback behavior, and tool restrictions. Documentation of available controls is not evidence they are enabled in someone’s deployment.
- Visibility and stopping
Control UI, run history, and an audit ledger expose activity, but the ledger is best effort with bounded retention and coverage gaps. Stopping a run, killing an already-running background process, and disabling/removing stored automations are separate operations. Automation history distinguishes execution from delivery; verify both before relying on an outcome.
OpenClaw: audit coverageOpenClaw: stopping workOpenClaw: managing automations
- Training policy
There is no single provider-independent training answer. Inspect the model provider, hosting arrangement, plugins, and services receiving data. OpenClaw’s mobile-app/extension privacy policy explicitly does not cover gateway or provider practices. Local hosting alone does not establish that prompts, embeddings, or tool results stay on the machine.
- Memory and deletion controls
The bundled memory-core supports forgetting tracked session-derived memories and excluding them from later ingestion. That does not erase original transcripts, every freeform memory edit, exports, backups, or other plugins’ stores. Identify the installed memory provider and all copies before treating a forget operation as complete deletion.
Meta MuseHosted personal agent · check current account and country eligibility
- Capabilities and practical fit
Meta documents inbox/calendar assistance, shopping, reservations, generated files, and custom tools using a persistent cloud computer/browser. These are useful personal-agent capabilities to consider alongside Dots, not measured success rates. Its advertised interaction surfaces include mobile, web, Mac, and WhatsApp.
- Availability and access permissions
Check current Muse account/country eligibility; this review did not establish a full rollout or pricing matrix. Some connectors support read-only access. Facebook, Instagram, and Threads connect automatically when they share Muse’s Accounts Center. Disconnecting stops further exchange but can leave information already in history or memory. Meta says custom connectors are not reviewed by Meta.
- Action approvals
Meta documents permissions by task, website, and connector, plus an external action-enforcement component called Sentinel. Certain actions require approval. Its September architecture article describes Confidential VM as forthcoming, not a current guarantee that Meta cannot access data. Operation, security, and support access remain relevant. These design claims do not establish comparative safety.
Meta: Muse guidance and approvalsMeta: Muse architecture and limits
- Visibility and stopping
An activity log records actions and permissions; Upcoming lists scheduled work. You can ask Muse to stop some actions, and some changes can be undone, but a sent email cannot be recalled this way. Review future schedules separately: the sources do not establish that stopping a current action cancels them all.
- Training policy
Training participation starts enabled. Meta’s Data controls let you disable it and say the choice also applies to previous interactions; this review does not establish reversal of completed training. Meta separately says Muse information is not shared with its ad systems. That advertising statement is not a promise of no model training or no service-provider access.
Meta: Muse data controlsMeta: Muse privacy and safetyMeta: Muse architecture and limits
- Memory and deletion controls
Memory files can be edited. Deleting a chat may leave information learned from it; requests to forget are described as best effort. Reset deletes chats, files, and active tasks. Check the deletion confirmation and any remaining schedules or connected-service data. The reviewed sources did not establish exact backup-deletion timing.
Four different questions about control
Instructions describe what you want. Access determines which accounts, apps, and files a tool can reach. Action approval decides whether a particular step can proceed. Data controls govern storage, memory, deletion, and permitted training uses. Changing one does not necessarily change the others.
For any assistant, check active work and future schedules separately. Stopping a run does not recall a sent message or undo a completed purchase. Turning off training is not the same as deleting stored data; disconnecting an app is not the same as forgetting information already obtained from it. Follow the specific product controls above.
Policies, incidents, and allegations are different evidence
A policy describes a vendor’s commitments. An incident report describes an event; a researcher’s demonstration establishes what happened in that tested configuration. An allegation remains an allegation until supported. None alone establishes which assistant is safest or most private.
A documented security advisory, with the maintainer’s response
The dated record below is separate from the current-policy comparison. It is not a comprehensive incident history or evidence of real-world exploitation. We do not infer that products without an example here have fewer incidents.
OpenClaw’s maintainer advisory GHSA-99qw-6mr3-36qr, published March 13, 2026, describes workspace plugins loading without an explicit trust/install step in versions through 2026.3.11. A malicious repository could consequently run code as the user. The maintainer’s stated response was version 2026.3.12, requiring explicit trusted state before execution, and a recommendation to upgrade. Read the original advisory and remediation.
This is a disclosed vulnerability with a reported fix, not an allegation of current compromise. We have not reproduced the issue or tested the patch. No other incident or allegation is asserted here without a verified primary record and a vendor response or an explicit account of its absence.
Choose one small, reviewable trial
Use the same non-sensitive notes and requested outcome in the tools you can access. Compare the result, corrections needed, visible actions, approval requests, and actual cost. Test stopping and deleting sample work before relying on an ongoing arrangement. These are suggested checks; we have not performed them for this guide.
No comparable hands-on evidence here supports an overall winner, or a claim that Dots is safer or more private. Start with the assistant whose documented access and controls fit your work, then verify the experience for yourself.
Try the illustrative project playbook · Learn how to steer the work
Official references
Source check: October 14, 2026. A source check is not hands-on testing.
- OpenAI: Create your dot in the ChatGPT mobile app — October 9 update
- OpenAI: Control your dot
- OpenAI: Dots privacy, security, and safety FAQs
- OpenAI: Tasks and memory
- Anthropic: Get started with Claude Cowork
- Google: Manage tasks and workflows in Gemini
- Microsoft: Copilot Cowork for personal accounts (Preview)
- OpenClaw: Documentation and overview
- Meta: Muse product overview