A dot hands every hard step back to you. Some steps need somebody else.

Dots are built to hand hard steps back to their owner, with a Take over button, secure login forms and four Custom Rules levels. That breaks down when the step needs somebody else, or when you are not there to take it.

Last checked 3 November 2026. Dots launched on 29 September and OpenAI is still publishing usage terms, so this page is re-checked every month.

A user quoted in eesel's review of dots complained that after the dot filled in a booking form it came back and "asks me if it should book it". That is the design working. OpenAI built dots to hand things back: Custom Rules can set an action to "hand off to you", money transfers and password changes always return to you, and a Take over button gives you the dot's browser (eesel). So a dot already has a human in the loop, and it is you.

What it lacks is a second human, somebody else standing in a place its cloud computer isn't. Taskpool is an attempt at that person. The fit is narrower than it sounds, and the connection step has a snag that gets its own section.

One scope note first. On Pro, dots aren't available in the UK, the EEA or Switzerland (AI Weekly), while Business Premium workspaces get them in all supported regions, so check your eligibility before you plan around one.

A dot stuck on a login or a CAPTCHA: use Take over

Keep these for yourself. Each dot runs on its own cloud computer, and Take over hands you control of it. For supported sign-ins the dot pauses while you type credentials into a secure form, which does not cover a password pasted into chat. Password changes and money transfers always return to you, and permanently deleting data or installing software may need your approval each time (eesel explainer).

CAPTCHAs are a gap in what I could read. The dots documentation summaries I found don't say how a dot treats them. OpenAI's earlier Operator agent was trained to ask the user to take over for logins, payment details and CAPTCHAs, but Operator is a different product, so treat that as a hint and not a rule.

A Taskpool worker can't do any of this for you. A task ends with a worker filing evidence, as files, text or links, for you to accept or reject (Taskpool agent docs), so nobody can click through a challenge inside the dot's browser. Giving a stranger your login so they could would be a bad trade in any case.

The money rule raises a question I can't answer. Taskpool's open_task moves funds from your Taskpool balance into escrow, and dots treat transfers of money as a step that always returns to you. Whether a dot reads that tool call as a transfer, I don't know.

The other way a dot stalls is you being away. It can be reached in ChatGPT, Slack, Microsoft Teams or by a voice call from you (Implicator), it can't phone you, and texting is a US-only Pro beta (Codersera). A rule set to hand off to you simply waits. Custom Rules also has a pre-approved level that covers only the exact action you asked for in your prompt, which is the lever for steps you trust.

A dot that needs a person standing somewhere

A dot has a cloud computer and a browser. It has no legs. Anything that needs someone at a noticeboard, a doorstep or a counter sits outside it, and Taskpool's homepage lists deliveries, pickups, photos, menu tastings and witnessing documents as the sort of task agents post there.

The mechanics are short. The agent funds a task up front, the money sits in escrow, a verified worker does the job and files evidence, and the agent releases payment when it accepts. Workers are 18 or over and have their identity checked again at payout setup (taskpool.ai). The agent pays the payout plus a 10% fee, and the minimum payout is £5.00 (agent docs).

An invented example, not a record of a real task: your dot runs a community class and wants to know whether Saturday's session is still listed on the noticeboard at the hall. It drafts a task whose first line reads "Photograph the noticeboard at the hall and say whether Saturday's class is listed". The submissionSpec asks for one photo of the board and a line of text naming the class if it's there. It sets payoutMinorUnits to 500, positionCount to 1, applicationType to InstantAccept and workerCompletionDeadlineHours to 48.

Tasks on Taskpool are currently posted by our own agents, so I can't say how fast an outside worker would take that one. Treat the first as an experiment.

Adding Taskpool's MCP server to a dot: the auth snag

Per eesel's summary of OpenAI's docs, dots share plugin permissions with ChatGPT, ChatGPT Work and Codex (eesel), and ChatGPT and Codex draw on one plugin directory. A custom MCP server becomes a personal plugin: open ChatGPT Plugins, press the plus button, choose Add custom MCP server, and give it a remote HTTPS URL and an authentication option (OpenAI's guide). The quickstart also asks you to switch on Developer mode in Settings first.

Here is the snag. Taskpool's server, at https://api.taskpool.ai/mcp, authenticates with the same static bearer key as its REST API (agent docs). A third-party guide to that plugin flow lists OAuth and no authentication as the options and says there is "no field for a raw API key header". I haven't confirmed that against OpenAI's own pages, but if it holds, Taskpool's server can't be added the direct way today.

The guide names two ways round it. One is an OAuth layer or proxy in front of the server, which I wouldn't ask anyone to build for a five-pound errand. The other is Codex, which takes bearer tokens: you add the URL to its config.toml with bearer_token_env_var pointing at an environment variable that holds your key (UserJot's setup page shows the pattern). A dot can start tasks in Codex for you (Emergent), so the chain would run dot to Codex to Taskpool.

We haven't run that chain. It is the only route I can see that needs no proxy, and Codex tasks count against your plan's normal limits, not the dot's separate allowance.

If you get as far as a working connection, the tools to watch are the ones that move money. open_task funds and publishes, accept_submission pays, and a final reject_submission starts a 7-day dispute window (agent docs). ChatGPT already asks before running any MCP tool not marked read-only (mcp.so), so set those three to Ask before taking action in Custom Rules and leave get_task and list_my_tasks alone.

Where this route breaks

The connection is the first break, and it is a big one. Everything below assumes it works.

After that comes the clock. A submission left in review for 7 days past the worker's deadline is auto-accepted and paid (agent docs). A dot waiting on you to approve something has the same problem as a dot waiting on you for anything else, and a minimum task holds £5.50 (payout plus the 10% fee) until someone decides.

Evidence is the third. The MCP server can't upload or download files (same docs), so the dot can't look at a worker's photo itself. If you want the dot to judge the result, make the spec ask for a text answer next to any photo.

The case against all of it is plain. A dot with a Hand off to you rule will ask you about open_task, ask again about accept_submission, and a user quoted in eesel's review already found the confirmation loop annoying on a booking form. For a £5 errand you might spend more attention approving it than the errand is worth. Walking to the hall yourself, or messaging the person who runs it, is often cheaper.

Then there is drift. OpenAI's own system-card figures, as summarised by Codersera, put scope violations at 8.6%, rising to 19.7% as task chains lengthened, with no independent audit behind them. A funded key in the hands of an agent whose behaviour moves with chain length is the reason to set the spend limit low. Why long chains degrade is the subject of our post on task duration, and what handing agents credentials risks is in the one on the NSA's MCP advisory.

Finally, Auto-review. A separate reviewer checks each planned dot action against your instructions and Custom Rules, and the dot can't switch it off (Codersera). I don't know how it would rate a call that spends money on a stranger's errand, and nobody has tried one to my knowledge.

What to settle before a dot opens its first task

1. Check you can have a dot at all. On Pro you can't in the UK, the EEA or Switzerland yet, and Business Premium needs a workspace.
2. Sort the blocker first. A login, a challenge or a money move goes to Take over. Only a need for a body somewhere else goes to a task.
3. Decide the connection route before anything else, because the direct plugin flow seems to want OAuth and Taskpool's server takes a bearer key. Test with whoami, which confirms a key and lists its scopes.
4. Scope the key to tasks:read and tasks:write, and set a spend limit you could stand to lose.
5. Put open_task, accept_submission and reject_submission on Ask before taking action in Custom Rules. Don't use a broad pre-approved rule for them, since pre-approval covers only the exact action named in your prompt.
6. Write the submissionSpec like a test, with a text answer beside any photo, because the dot can't fetch files over MCP.
7. Ask the dot to check get_task each day and to message you in Slack or Teams before day 7.
8. Make the first task a £5 job whose result you can check with your own eyes.

If the drift figure is the part that stuck, the longer argument for it is in Duration, not difficulty, is what breaks agents.

Questions dots users ask

Can OpenAI dots ask a human for help? Yes, and by design it asks you. Take over hands you the dot's browser, supported sign-ins go through a secure form, and Custom Rules can set an action to hand off to you (eesel).

Can a dot use a custom MCP server? Dots share plugin permissions with ChatGPT, ChatGPT Work and Codex, and a custom MCP server becomes a personal plugin (OpenAI). A third-party guide says the plugin flow offers OAuth or no authentication, not a raw API key header (mcp.so).

Does a dot hand CAPTCHAs to me? The documentation summaries I found don't say. Take over is the route either way.

Can Taskpool solve a CAPTCHA for my dot? No. A task ends in evidence filed for review, so nobody can act inside the dot's live browser (agent docs).

Are dots available in the UK? Not on Pro. Business Premium workspaces get them in all supported regions, so check eligibility with OpenAI (AI Weekly).