AgentComparison
Comma · product · personal assistant

Comma

A personal agent that works through ongoing tasks with a cloud computer.

Group: Errands, travel and food

Comma describes a personal agent organised around ongoing Tasks, with connected apps, Mac access and a cloud computer.

Documentation researched, not testedChecked 23 October 2026Vendor documentation, not a test

The decision in plain English

Distinguish cloud work from tasks needing Mac files or call recording; check the actual approval settings before connecting apps.

Best fit

Personal or team app tasks that continue in a cloud environment.

Poor fit

A user who needs every connected-app action to wait for approval without configuring boundaries.

Documented capabilities

Vendor statements, not verified task outcomes.

  • Tasks that continue on a cloud computer Source · checked 2026-10-23
  • Connected app and file work Source · checked 2026-10-23
  • Routines and proactive messages Source · checked 2026-10-23
  • Signal, Telegram and WeChat access Source · checked 2026-10-23
  • Router/Worker model: one Router hands each Task to a Worker; choose a model per role or let Codex or Claude Code be the Worker Source · checked 2026-10-23

Permissions and human control

  • Connected apps get scoped access per the data-use page.
  • Always-on, proactive, sessionless agent: one agent keeps work moving across Tasks on a board (In progress / Needs review / Done).
  • BYO model option: use your own API key, or your ChatGPT or Claude plan - those calls use no Comma credits.
  • The cloud computer consumes credits while it runs; model calls are charged at the model's published API price in credits.

Documented controls do not prove that an agent always follows them.

Limits and open questions

  • Call recording and Mac-local files need the Mac
  • Landing-page demonstrations are not verified task outcomes

Privacy and security

Comma's data-use page states conversations, prompts and files are never used for training and personal data is not sold; user-submitted content is not used to train its own or third-party general-purpose AI models, except where a customer expressly requests a separate training or fine-tuning workflow under specific terms. Connected apps get scoped access. A published retention table covers the main categories: workspace content and saved workflow context retained until deleted; intermediate outputs typically deleted within 30 days unless promoted to persistent state; recordings and uploaded audio retained until deleted. Checked 23 October 2026.

Feature-by-feature: what is documented and what is not

Official sources checked 23 October 2026. A documented feature is a vendor statement, not proof of success.

FeatureEvidence statusDetail and limit
PricingFree $0; Pro $80/mo or $800/yrRecord verified 7-22 Oct: free tier $0/month; Pro $80/month or $800/year billed annually ('Save $160'). Monthly credits expire after one month; credit packs last to end of the purchase month. Source
Training useNever, per data-use pageConversations, prompts and files never used for training; no training of own or third-party general-purpose models on user content; exception only for express customer-requested training workflows. Source
Credit modelAPI-price passthroughModel calls cost the model's published API price in credits (e.g. a $2/M-token model uses 2 credits per input token); BYO key/ChatGPT/Claude plan uses no credits; cloud computer consumes credits while running. Source
RetentionPublished tableWorkspace content until deleted; intermediate outputs typically 30 days; recordings/audio until deleted; configurable. Source

How to start with Comma narrowly

  1. Start on the free tier and run one Task end-to-end (In progress to Done) to see the board model before paying.
  2. Check Settings > Usage & billing for the credit renewal date - monthly credits expire after one month.
  3. If you already pay for ChatGPT or Claude, connect your own key or plan so model calls skip Comma credits.

Official product/setup source. These are documentation-based steps, not a hands-on setup test.

Common questions about Comma

What is Comma for?

Personal or team app tasks that continue in a cloud environment. The vendor describes it as: Comma describes a personal agent organised around ongoing Tasks, with connected apps, Mac access and a cloud computer.

Who should look elsewhere?

A user who needs every connected-app action to wait for approval without configuring boundaries.

How much does Comma cost?

Value $0/month. Pricing toggle checked 7 October shows Pro $80/month and $800/year with Save $160; Billed yearly label persists in both states, so confirm final checkout period. About 300-600 tasks/month is a vendor estimate. 1M credits = US$1; cloud computer about 74,000/hour while running; model calls at API cost. BYOK model calls use no Comma credits, but compute still does. Credits do not roll over. Re-checked 22 October 2026 on comma.surf (the /pricing path now redirects to the homepage pricing section): Free $0 BYOK and Pro $80/month billed yearly (save $160) unchanged. Monthly Pro $80 re-verified 22 October 2026; the annual $800 figure is from the 7 October toggle read and was not re-verifiable in the 22 October text pass. New from the 22 October FAQ: 1M credits = US$1 of usage at published API prices; the cloud computer uses about 74,000 credits per running hour, billed by the second and shut down ~5 minutes after last work; plan credits expire monthly.

What can it access and what needs my approval?

Connected apps get scoped access per the data-use page. Always-on, proactive, sessionless agent: one agent keeps work moving across Tasks on a board (In progress / Needs review / Done). BYO model option: use your own API key, or your ChatGPT or Claude plan - those calls use no Comma credits. The cloud computer consumes credits while it runs; model calls are charged at the model's published API price in credits.

What happens to my data?

Comma's data-use page states conversations, prompts and files are never used for training and personal data is not sold; user-submitted content is not used to train its own or third-party general-purpose AI models, except where a customer expressly requests a separate training or fine-tuning workflow under specific terms. Connected apps get scoped access. A published retention table covers the main categories: workspace content and saved workflow context retained until deleted; intermediate outputs typically deleted within 30 days unless promoted to persistent state; recordings and uploaded audio retained until deleted. Checked 23 October 2026.

Has AgentComparison tested it?

No. This record is built from the vendor's own documentation, checked 2026-10-23. Vendor claims are attributed, not confirmed.

How to evaluate Comma yourself

  1. Pick one low-risk task that matches its documented capabilities: tasks that continue on a cloud computer.
  2. Write the expected output and what counts as failure before you start.
  3. Connect the minimum access needed, and check approval settings first.
  4. Run it twice and compare, then review every action it took.
  5. Check cost after any free allowance. Unknown price is not free.

Related

Documented use cases: Real-world errands and long-running tasks.

Sources and history

Change history

Documentation overview expanded with current pricing/access sources. No hands-on results claimed.

Report a correction