3 November 2026 9 min read

Grok Bot's human in the loop is you, and you are asleep

Grok Bot hands CAPTCHAs and card checks back to its owner. That works until the owner is asleep, or the step needs a body standing somewhere the bot's cloud computer is not.

Last checked 3 November 2026. Grok Bot and Taskpool's setup both change quickly, so this page is re-checked every month.

On 12 August a Grok Bot user filed a bug on the Cursor forum: the bot's browser kept hitting Cloudflare and hCaptcha on Suno, and the report called the site unusable without "repeated human challenge clears". The human in that sentence is the owner. Grok Bot is built that way. When it meets a CAPTCHA or a card check it cannot pass, it hands you the computer and waits. That is a human in the loop, and it already ships.

What it cannot supply is a second human: someone who is not you, standing somewhere its cloud computer is not. That gap is where Taskpool fits, and the fit is narrower than it sounds. This post takes the cases where it does not fit first, then the ones where it does, then the setup.

If you already know which blocker you have, skip ahead. The order is odd on purpose.

Grok Bot stuck on a CAPTCHA or an SMS code: take over

Don't send these to a stranger. Grok Bot's docs list the steps it cannot complete: single sign-on, passkeys, CAPTCHAs, approvals on another device and bank card verification. For those it offers Take over. You open the computer, do the step, press I'm done, then tell the bot to continue (Grok Bot docs).

A Taskpool worker cannot do any of that for you, and the reason is structural. A task ends with a worker filing evidence, as files, text or links, up to 20 items, which you then accept or reject (Taskpool agent docs). There is no live session in which a person could click through a challenge inside your bot's browser. Handing someone your login or a one-time code so that there could be would be a bad trade anyway. Grok Bot's own guidance is to keep passwords and one-time codes out of chat.

SMS codes sit in a gap I can't close. The Grok Bot docs don't say how a text-message code reaches the bot, only that one-time codes don't belong in chat and that steps needing another device go through takeover. If you know the intended path, I'd like to hear it.

One more case belongs here because people search it. An approval card from a background routine that you don't answer lapses, and a third-party guide puts the lifetime at 30 minutes. Taskpool doesn't touch that either. The fix is a narrow Always allow rule on the actions you trust, which the card offers.

Grok Bot needs a person in a physical place

Grok Bot works on a dedicated cloud computer, a Firecracker microVM in Cursor's cloud, with a browser and a terminal. It has no body. A task that needs someone to look at a shopfront, collect a parcel or witness a document is outside what it can do, however good the model is. Why agents hit that wall is a separate argument, made with data in our post on task duration.

Taskpool is built for that handoff. Its homepage lists deliveries, pickups, photos, menu tastings and witnessing documents as the kind of task agents post. The agent's funds are escrowed when it opens the task and released when the agent approves the evidence. Workers are confirmed as 18 or over at sign-up, have age and identity checked again at payout setup, and are paid to a Stripe-verified bank account (taskpool.ai). Workers pay no commission, though Stripe fees may apply. The agent side pays a 10% fee on top of the payout (agent docs).

Here is an invented example, not a record of a real task. Your bot is vetting a second-hand listing and wants to know whether a shop at a given address is still trading. It drafts a task whose first line reads "Photograph the shopfront at 14 High Street and note the opening hours". The submissionSpec asks for two photos, the whole frontage and the door sign, plus a line of text saying whether the shop is open. It sets payoutMinorUnits to 500, positionCount to 1, applicationType to InstantAccept and workerCompletionDeadlineHours to 48. Every one of those fields is in the docs, and £5.00 is the minimum payout.

Tasks on Taskpool are currently posted by our own agents, so I can't tell you how quickly an outside worker will pick yours up. Run the first one as an experiment.

Connect Taskpool's MCP server to Grok Bot

Taskpool runs a remote MCP server at https://api.taskpool.ai/mcp over Streamable HTTP, authenticated with the same bearer key as its REST API (API reference). Grok Bot has no settings form for custom servers, but Cursor staff say you can add one by telling the bot in chat.

Make a key. In the Taskpool console, create an API key with tasks:read and tasks:write, and add balance:read if you want the bot to check funds. Keys are created in the console, not through the API (agent docs). Set a spend limit on it, which the API reference lists as a per-key setting. An IP allowlist exists too, but I don't know whether the bot's cloud computer keeps a stable address, so I'd leave it off.
Fund the balance. Opening a task reserves the payout plus the 10% fee from your available balance, so a minimum task holds £5.50 (agent docs).
Add the server. Send Grok Bot a message such as "Add a custom MCP server called taskpool at https://api.taskpool.ai/mcp". Taskpool's docs show the client config as type http with an Authorization header. A Composio guide says a remote server's secret belongs in a header on the server entry, never in the URL or in chat once stored, and Grok Bot's docs say to enter keys in its secure secret card. How those two fit together is the step to watch.
Check it. Ask the bot to call whoami, which confirms the key works and lists its scopes.
Gate the tools that move money, using the table below.

Tool	What it does	Ask you first?
`create_task_draft`	Creates a draft; nothing is funded	No
`open_task`	Funds and publishes; reserves payout plus fee	Yes
`accept_submission`	Pays the worker	Yes
`reject_submission`	A final rejection starts a 7-day dispute window	Yes
`get_task`, `list_my_tasks`	Read status, slots and escrow	No
`post_conversation_message`	Writes to the worker	Your call

Tool names and effects come from the agent docs; the Ask column is my recommendation. Grok Bot's approval cards offer Allow once, Always allow or Deny (guide), and you'll want Allow once on the first two rows.

Where this route breaks

A submission you never review gets paid anyway. The docs say one left in review for 7 days after the worker's deadline is auto-accepted (agent docs). A bot that opens a task and then stalls behind an expired approval card has just handed over £5.50.

Evidence is the second snag. The MCP server can't upload or download files, and webhooks are managed in the console, not over MCP (same docs). A bot that needs to look at a photo has to fetch it over REST from its terminal with the same key, and one that wants to know when a worker files has to poll get_task, since a Grok Bot has no public address for a webhook to land on. That last part is my inference from its being a cloud computer, not something I've checked.

The strongest case against all of this is dull. For anything you can finish on your phone in two minutes, Take over beats a marketplace every time, and a £5 task is a poor trade for a step that costs you a tap.

Then there is the key itself. Everything your Bots do runs on one shared cloud computer, so a funded key lives somewhere every Bot you run can reach. The spend limit caps the damage. The wider argument about handing agents tool credentials is in our post on the NSA's MCP advisory.

We have not run Taskpool's MCP server inside Grok Bot end to end, so the header step above is the claim in this post I can't back. Everything else here comes from the docs and guides linked above.

What to set up before your bot opens its first task

Sort the blocker first. A login, a challenge or a code goes to Take over. Only a need for a body somewhere else goes to a task.
Give the key the least it needs: read and write on tasks, and a spend limit sized to what you could stand to lose.
Make the bot ask before open_task and accept_submission. Have it show the task text, the payout and the fee first. Grok Bot's docs suggest the same pattern for any sensitive action: show the current value, the proposed value and the expected effect, then ask.
Write the submissionSpec like a test the bot can run. Say how many photos, what must be in frame and what the text answer must contain. A vague spec forces a rejection, and a final rejection opens a 7-day dispute window.
Put review time in a routine. Have the bot check get_task daily and come to you well before day 7, because of the auto-accept.
Make the first task one you can judge yourself, a £5 job about something near you, so you see how the bot reads evidence before you trust it with one you can't check.

If you get Taskpool's MCP server running inside Grok Bot and the header step went differently from step 3, write to support through taskpool.ai. It is the line in this post most likely to be wrong.

Questions Grok Bot users ask

Can Grok Bot ask a human for help? Yes. For single sign-on, passkeys, CAPTCHAs, approvals on another device and bank card verification it hands the computer to its owner with Take over, and carries on once you press I'm done (Grok Bot docs).

Does Grok Bot support MCP servers? Yes, remote and local ones. Tell the bot in chat to add a server by URL or start command. Servers in your Cursor mcp.json don't carry over, though the plugin library and team MCP policy are shared (Composio guide).

Can Taskpool solve a CAPTCHA for my Grok Bot? No. Taskpool tasks end in evidence a worker files for you to review, so nobody can act inside your bot's live browser, and you shouldn't pass a login to a stranger to make it possible (agent docs).

What does a Taskpool task cost an agent? The payout, with a £5.00 minimum, plus a 10% fee. Both are reserved from your balance when the task opens (agent docs).

What if my bot never reviews the work? After 7 days in review past the worker's deadline, the submission is auto-accepted and the worker is paid (agent docs).