# Cursor’s Cloud Agents Take Coding Jobs From Grok Bot

**Grok Bot can hand coding work to Cursor Cloud Agents through a four-step contract: reproduce the bug, request a fix, test it, and open a draft PR.** Cursor’s [team controls](https://cursor.com/docs/grok-bot/teams), checked October 12, 2026, confirm that Bots can delegate coding tasks and admins can disable delegation. A draft pull request proposes a change; it does not authorize a merge.

Grok Bot is an agent app with persistent roles, context and a hosted computer. The useful unit of work is a reviewed artifact with evidence attached. Naming a Bot “Repo Maintainer” gives the work an owner. The instructions below give that owner something checkable to deliver.

These ten steps keep the workflow in order. Seven copyable text templates describe the requested behavior, subject to supported integrations, permissions and approvals. They are instructions, rather than Python programs or measured results.

![Grok Bot's official product page showing named Bots and a complete conversation interface.](https://s4.tenten.co/learning/content/images/2026/10/landing-page-1-4.png)

*[Official product demonstration](https://x.ai/bot). The Sales Outbound draft counts are demo content, not results from this article's implementation.*

#### 01. Backend upgrades leave model selection with the product

Grok Bot manages model selection and provides no model picker. Elon Musk’s [October 7, 2026 announcement](https://x.com/elonmusk/status/2107724314451878104) says SpaceX will use the best backend model for each task, naming Claude Opus 5.5, Midjourney and Suno.

The announcement does not assign a release date to every integration or promise that every Bot action uses Opus 5.5. Adding a model name to a charter does not create a selection control.

Anthropic’s API specification gives Opus 5.5 a one-million-token context window and a 128,000-token maximum output. A token is a unit of text processing. Those API limits do not establish Grok Bot’s per-task capacity. Judge the workflow by its artifacts and checks, rather than multiplying a general benchmark score into a promised success rate.

#### 02. Install the app and give the first Bot one job

Grok Bot supports macOS, Windows and Linux desktop clients, plus iPhone, iPad and Android. Users sign in with a Cursor account and link an eligible individual SuperGrok subscription when needed.

Start with a coordinating Chief and one bounded assignment, such as a source-linked inventory of a specified folder. Use New in the sidebar, then Create new Bot, and set a name and job. An inspectable output is the signal to add another role. A reassuring chat response is too little evidence.

#### 03. The Research Lead owns verification and the saved draft

A Research Lead needs an artifact, acceptance criteria and a stopping point. Grok Bot’s durable context makes a precise charter reusable across later assignments.

```text
Research Lead charter for tasks handled by Opus 5.5 or another backend
You are my Research Lead.

// Responsibility
For each topic, complete the research assignment: locate primary sources,
open every cited page, check figures against at least two sources,
and produce a 1,500-word brief.

// Acceptance criteria
Put a source link beside each number. Mark conflicting accounts;
do not average their figures. Save the draft in my Drive folder.
A chat-only response does not satisfy the delivery requirement.

// Stop conditions
Drafts only. Do not publish or email anyone.
If conflicting sources remain unresolved, pause and ask me.
```

A brief without citations fails this contract even if its prose reads well. A brief that never reaches Drive fails a different part of the same contract. These checks test the work the Bot was hired to do.

#### 04. Connected accounts belong to the shared computer

One user’s Grok Bots share a hosted computer, including files, browser sessions and app logins. Separate conversations do not give separate roles independent credentials.

Install needed plugins through Marketplace and grant only the access the assignment requires. A role description cannot isolate legal files from a browser session available to the same user’s other Bots.

```text
// Data boundary: include this in every charter
Do not open, read or summarize:
  - content in my /Legal or /Finance Drive folders
  - email from my lawyer, accountant or bank
  - messages containing passwords, seed phrases or 2FA codes

Pause and ask me if the task requires any of those inputs.
Use only the portion of a document needed for the assignment.
```

Cursor’s security documentation describes requests passing through its servers to model providers. With Privacy Mode enabled, customer data is not used for training. Provider retention agreements include exceptions for abuse investigations. An outside model’s inclusion does not establish one universal provider policy for every task.

#### 05. The user completes sensitive sign-in steps

Grok Bot should hand over computer control for passwords, two-factor checks, CAPTCHAs and payment confirmations. Open Agent Computer, take control, complete the sensitive step, then return control.

Keep passwords and one-time codes out of ordinary chat. Supported connections also offer a masked secure secret request that keeps the value outside the transcript and away from the model. Confirm the service and account before using it. Once a browser login persists, the same user’s other Bots may use that session.

#### 06. A successful task becomes a tested routine

Grok Bot skills store reusable instructions; routines repeat work on a schedule or supported event. Teach-by-demonstration is rolling out gradually, so a missing Teach control calls for a written skill built from a completed task.

![Four Grok Bot capability cards covering sign-in, demonstration, retained context and connecting Bots.](https://s4.tenten.co/learning/content/images/2026/10/landing-page-2-4.png)

*[The maker's feature overview](https://x.ai/bot); account permissions, Teach rollout and routine acceptance still require separate checks.*

Demonstrations can record up to ten minutes and do not record microphone audio. Add failure handling and approvals before saving a skill. Confirm the schedule’s time zone; the times below are example settings.

```text
// Schedule: morning brief
At 7am each weekday, review my calendar, inbox and #launches Slack channel.
Produce a short brief covering today's commitments, replies needed,
and changes since yesterday evening.

// Event requirement: new pricing inquiries
When email from a domain outside my contacts mentions pricing,
prepare a reply using the template and hold the draft for me.

// After an assignment produced the result you wanted
Run this assignment every week.
```

The email listener is a requirement to check, rather than a capability unlocked by installing a Gmail plugin. Documented event integrations include Slack messages and GitHub notifications connected through the Cursor account, separately from plugins. Use a supported scheduled check if the email event is unavailable. Review the next run and test with safe inputs. Test run performs real operations.

#### 07. The Repo Maintainer hands off code and collects proof

Grok Bot can delegate to Cursor Cloud Agents running on separate development computers. Team admins control delegation, and the Cloud Agent environment and network policy are separate from the Bot’s shared computer.

Configure the repository, dependencies and test environment first. A Cloud Agent that cannot install a private package or reach a required service cannot establish that its fix passes the full suite.

```text
// Repo Maintainer: a Bot that supervises coding agents
You are my Repo Maintainer.

When a GitHub issue receives the "bug" label:
1. Reproduce the problem on your computer and add a failing test.
2. Start a Cursor cloud agent to fix the issue on a new branch.
3. Run the full test suite against the returned change.
4. Open a draft PR containing the test, fix and a 3-line summary.

Do not merge. Do not push to main.
Stop for my review if the fix touches authentication, billing or a migration.
```

This is a delivery contract, conditional on issue matching, repository access and permitted delegation. If spawning or testing fails, the Repo Maintainer should identify the blocked step. An attractive draft PR cannot compensate for missing execution evidence.

For a sample issue, require reproduction instructions, the test failure before the fix, the passing result afterward and the complete suite output. Those are proposed acceptance fields, not vendor-reported results. Stop before sensitive changes. Use repository permissions and branch protections to enforce the merge boundary as well as stating it in chat.

#### 08. Group coordination keeps artifact ownership visible

Grok Bot supports Bot-to-Bot messages, shared group context and task handoffs. A coordinating Chief can track dependencies while each specialist remains responsible for its own deliverable.

```text
// Set an outcome and let the roles divide the work
Goal: launch the October newsletter by Friday 5pm.

Research Lead owns facts and source checks.
Inbox Manager owns subscriber replies.
Chief owns the schedule and reports blockers to me.

Divide the work among yourselves. Hold all public actions for my approval.
```

The deadline and approval boundary apply together. Research can deliver a checked draft, and Inbox Manager can prepare replies. Chief should report missing work or approval; coordination does not itself authorize publication.

#### 09. Approval rules constrain even reversible work

Grok Bot’s Auto Review evaluates actions before execution. Personal rules sit under Settings, General, Auto-review; when Ask first and Allow automatically both match, Ask first takes precedence.

```text
// Complete these within the approval rules and log the work
draft, file, tag, summarize, research, prepare, reconcile
Keep these operations reversible. Proceed without another question
only when the applicable approval rules permit it.

// Hold these for me
send anything to someone outside the company
spend or move money, or commit to a price
publish anything publicly
delete anything other than obvious junk
accept terms or register for a service

// If the effect is uncertain
Pause and ask if you cannot undo the operation in under a minute.
```

The template records preferences; actual permissions still matter. Put deletion behind Ask first even when an item appears to be junk. External email, payments and production changes also need identifiable targets and explicit approval.

Approval requests from a routine, event or another Bot expire after roughly ten minutes, and the action does not run. An interactive conversation waits for the user’s answer. Distinguish executed work from pending and expired approvals in the run log.

#### 10. Weekly allowances and API prices answer different questions

Grok Bot’s [plans and billing](https://cursor.com/help/grok-bot/plans) specify weekly usage resets. After the allowance runs out, enabled on-demand usage continues through Cursor billing and counts toward a monthly limit.

Linked SuperGrok and Cursor allowances do not stack; the larger eligible grant applies. A monthly limit may let an already-running Bot finish beyond the limit. Turning off on-demand stops additional work after the weekly allowance is exhausted until reset. Inspect Usage & Billing rather than assuming a flat subscription guarantees unlimited work.

![Grok Bot's official plan cards showing monthly and per-seat units and access through eligible subscriptions.](https://s4.tenten.co/learning/content/images/2026/10/landing-page-3-4.png)

*[Official page state](https://x.ai/bot) on October 12, 2026. The article's standalone API rate comparison is separate from these subscriptions.*

Standalone API rates offer a model-cost comparison. They do not determine the app’s bill. These standard, uncached global rates were checked October 12, 2026; tools and other premiums are excluded.

| Model and prompt size | Input per million tokens | Output per million tokens | Context limit |
|---|---:|---:|---:|
| Claude Opus 5.5 | $4 | $20 | 1,000,000 tokens |
| Grok 4.7, below 200,000 input tokens | $2 | $6 | 500,000 tokens |
| Grok 4.7, at least 200,000 input tokens | $4 | $12 | 500,000 tokens |

Anthropic’s [published rates](https://platform.claude.com/docs/en/about-claude/pricing) and xAI’s [API rate card](https://docs.x.ai/developers/pricing) establish the comparison. Grok 4.6 has the same listed rates. The long-context Grok rate applies to the entire qualifying request. Equal $4 input rates and $12 versus $20 output rates do not make Opus cheaper after the threshold.

For an illustrative request with 30,000 input and 8,000 output tokens, the calculated token cost is $0.28 for Opus 5.5 and $0.108 for short-context Grok 4.7. Retries and tool charges are excluded. Check the product meter to learn what a Grok Bot assignment actually consumes.

```text
Report every routine you ran this week. For each routine include:

  - the number of times it triggered
  - the artifact it produced
  - skipped work, failures and unresolved guesses
  - items waiting for my approval that I never answered

Identify the least useful routine and explain the choice.
Say when you cannot determine which model handled a step.
```

Compare the weekly log with Usage & Billing. Pause routines that produce no usable artifact, repeatedly expire awaiting approval or depend on an unavailable source. Adding another coordinator would leave those problems intact.

#### FAQ

##### Can Grok Bot setup include handing bugs to Cursor Cloud Agents?

Grok Bot can delegate coding work when the account and team permit it and the Cloud Agent has the required repository and test environment. The four-step template ends at a draft PR and does not authorize merging. Authentication, billing and migration changes require human review under that template.

##### Does Grok Bot pricing include unlimited Opus 5.5 use?

Grok Bot paid access includes a weekly usage allowance, not unlimited model usage. Enabled on-demand use after that allowance is exhausted is billed through Cursor. Grok Bot has no model picker, and a standalone API price does not establish its per-task app charge.

##### Does the Grok Bot free trial guarantee seven unlimited days?

Grok Bot’s free trial is a usage credit with a seven-day window. Agent steps and tokens consume the credit, so a long assignment can exhaust it early. An iOS App Store introductory offer is separate and follows the store’s subscription and cancellation rules.

##### Which Grok Bot download matches a desktop computer?

Grok Bot has Apple silicon and Intel macOS builds, plus x64 and Arm64 builds for Windows and Linux. Check the machine’s architecture before choosing an installer. Linux downloads include `.deb`, `.rpm` and AppImage formats.

##### What are Grok Bot Android and iPhone requirements?

Grok Bot requires Android 9 or later on Android and iOS 18 or later on iPhone, according to Cursor’s deployment documentation. Installing a mobile client does not add another Grok Bot usage allowance. Access remains tied to the signed-in account.

#### Sources

- [Grok Bot official product page and captured media](https://x.ai/bot)
- [Team administration and coding delegation](https://cursor.com/docs/grok-bot/teams)
- [October 7, 2026 backend-model announcement](https://x.com/elonmusk/status/2107724314451878104)
- [Model selection and settings](https://docs.x.ai/grok-bot/settings-and-notifications)
- [Product overview and shared computers](https://docs.x.ai/grok-bot/overview)
- [Creating and managing Bots](https://docs.x.ai/grok-bot/bots)
- [Apps, plugins and sign-in handoffs](https://docs.x.ai/grok-bot/computer-and-apps)
- [Skills, routines and event integrations](https://docs.x.ai/grok-bot/skills-routines-and-automations)
- [Approval rules and expiry](https://docs.x.ai/grok-bot/approvals-security-and-privacy)
- [Data handling and security architecture](https://cursor.com/docs/grok-bot/security)
- [Cloud Agent environment setup](https://cursor.com/docs/cloud-agent/setup)
- [Plans and billing](https://cursor.com/help/grok-bot/plans)
- [Desktop and mobile deployment requirements](https://cursor.com/docs/grok-bot/deployment)
- [Opus 5.5 API specifications](https://platform.claude.com/docs/en/models/opus-5-5/overview)
- [Anthropic API pricing](https://platform.claude.com/docs/en/about-claude/pricing)
- [xAI API pricing](https://docs.x.ai/developers/pricing)

#### Author Insight

Teams unable to answer Grok Bot’s autonomous approval requests within roughly ten minutes should keep approval-heavy repair work interactive. A staffed conversation lets a Repo Maintainer present the draft and wait for a decision. Approval expiry is a documented product condition; switching the backend model cannot supply the missing operator.

#### Terminology

| Term | Meaning in this workflow |
|---|---|
| Bot | A persistent role with its own conversation and working context |
| Cloud Agent | A coding agent on a separate development computer |
| Skill | Reusable work instructions |
| Routine | Work started by a schedule or supported event |
| Auto Review | Model-based action review before execution |
| Draft PR | A proposed code change awaiting review and merge |
| Token | A unit of model input, output and billing |
| On-demand | Additional billed use after the weekly allowance |