Maya: Welcome to the Cursor HQ daily briefing. Hi, I'm Maya. We're tracking where Cursor's agents actually run, a quiet Sonnet charge, and two free-access offers worth checking before you act.

James: Hi, I'm James. Let's start with a big shift in how Cursor's Cloud Agents work. On September 2, Cursor added a Remote Machines option that runs agent code on servers you control.

Maya: Right. Cursor still handles the AI reasoning and planning, but execution can now sit inside your own network. Supported targets include AWS Lambda, Vercel, Cloudflare, E2B, Modal, Daytona, Coder, Namespace, and self-hosted machine pools.

James: The consequence is access. Agents can now reach internal databases, private Git repos, or GPUs without handing that access to an outside sandbox. Cursor says Cloud Agents already generate over 60% of its merged pull requests.

Maya: Providers moved fast. Cloudflare confirmed support the next day, Modal positioned itself for scaling to millions of parallel workers, and on September 7 Cua added Windows and Linux VMs, not just macOS.

James: One caveat for teams: you pay the compute provider directly, and Cursor isn't reselling that. Running your own auto-scaling pool is real infrastructure work the settings panel doesn't erase. Speaking of quiet costs, our next story.

Maya: A Reddit report from user vibesresearcher says Cursor's Cloud Agent runs a browser helper that always calls Claude Sonnet, no matter which parent model you picked for the session.

James: And it's hard to spot. The post says that usage logs under the generic Other Models line, and it draws down tokens even when you've deselected Sonnet everywhere else in settings.

Maya: For Ultra subscribers who pick cheaper models like Grok or Luna to stretch usage, that's a quiet leak in the budget. The whole point was predictable spend, and this undercuts the math.

James: There's no setting to redirect the helper, so it's fixed behavior. The practical advice: check your usage dashboard for unexpected Other Models charges, and treat any browser task as Sonnet-backed. Now to a free offer.

Maya: A TikTok post from creator Krix.sync on September 3 says SuperGrok Heavy subscribers can claim one free month of Cursor Ultra, the forty-dollar tier with unlimited fast requests and premium models.

James: No credit card, no auto-renewal. The trial just lapses after 30 days. Since Ultra sits above Pro on rate limits and model access, it's a clean way to test whether unlimited fast requests actually change your day-to-day workflow.

Maya: The claim mechanism is unusual, though. Instead of a checkout code, you comment ULTRA on Krix.sync's video, and that comment triggers access with no payment method at any point.

James: Worth flagging: the source is one creator's TikTok, not Cursor or xAI. There's no changelog or official partnership confirming it, so verify it lands in your account. The deal expires permanently after one month, with no grace period or renewal path, so convert before the clock runs out. Next, a bigger giveaway.

Maya: The X account Teslaconomics posted on August 30 that Grok Bot giveaway codes now redeem for a confirmed 200 dollars in Cursor value, up from 60. That covers a full month of Cursor Ultra or matching API credit.

James: The timing is notable. Grok Bot's own entry price has been falling, opening to Cursor Pro at twenty dollars on August 26. Yet the code value tripled to match the ceiling, not the floor.

Maya: Grok Bot is Cursor's autonomous background agent built on Grok. It draws its own weekly allowance, so handing it work doesn't eat your regular quota, and it authenticates through Cursor accounts.

James: To claim one of 50 codes, you reply describing a job you'd assign, and winners are picked from the replies. One piece of fine print: linking a SuperGrok or X Premium+ account to Grok Bot is permanent and can't be moved later. Our last story tackles quality.

Maya: Developer Manish Kumar reported on X that Antislop AI shipped a 38-rule filter, indexed R-01 through R-38, that runs inside Cursor, Claude Code, and Codex to catch generic layouts, placeholder copy, and redundant comments before they land.

James: It installs with one command, npx antislop-ai, no configuration needed. Rules split into three tiers: hard gates that are absolute, purpose gates that require documented reasons, and quality locks that check consistency across a build.

Maya: The key point is it isn't a style guide. The README calls it a filter with no prescribed colors or fonts. It pairs with your own DESIGN.md, so your existing style reference becomes the source of truth. It runs during generation and audits finished work afterward.

James: It's MIT-licensed, currently at 817 stars, and it's not alone. A separate tool called Hallmark runs 58 deterministic gates. The consequence is that a rule-based check runs before frontends ship looking interchangeable, instead of asking the same model to review its own output.

Maya: The through-line: more control over where agents run and what they produce, but more responsibility on you to watch costs and verify offers.

James: Check your usage dashboard, confirm any free access lands, and read the fine print before committing. That's your Cursor HQ daily briefing.