Talk to a Cursor agent in a realtime voice call, plus ChatGPT dots and sign-in codes.

The Daily Bot

No. 027 · Monday, Nov 9, 2026

Note: The Daily Bot website is temporarily offline after hitting its hosting transfer limit. Until it is back, this email carries every story in full. Source links below still go to the original posts, docs and changelogs.

Today's one job is to try a realtime voice call with a Cursor agent, then decide whether your ChatGPT dot should be allowed to fill emailed sign-in codes. Also inside: a Grok Bot sales ops playbook, how users are moving Grok bot jobs into Hark, a Cursor tip to refuse shell commands outside the sandbox, Friday's week in review, and the full Monday digest with every known issue.

In this issue: Week in review · Cursor voice chat · What's new (1) · How people use it (3) · Digest · Known issues · Coming up · Daily Bot Check

The week in review

Grok Bot this week: phone calls official, Cursor live apps (Grok Bot · reported)

Nov 2 to 6: official phone calls, Cursor live apps and Voice Chat, Claude Personal Finance, Images 2.6, what broke, and what to watch next.

Try phone calls only with a hard stop list, then decide whether a live Cursor app is worth publishing for your team. Outside calls are now official from Grok Bot's own account, Cursor shipped Build Live Apps and announced Voice Chat with Agents, Claude opened Personal Finance in the US, and OpenAI shipped Images 2.6 with GIFs. The Grok Bot changelog is still on v0.78.1 from Oct 27. Two short Friday incidents came and went.

The short version

• New in Grok Bot: official outside phone calls (confirmed by @bot; still absent from the changelog), a weekly usage meter scare, and a shared Data Bot playbook from SpaceXSI's own team.
• New in Cursor: Build Live Apps (Nov 5 changelog), Voice Chat with Agents (Nov 6 changelog; our story dates Monday).
• Broke: a Nov 1 billing/usage dashboard degradation, Friday computer-use regional and Grok Bot degraded windows (both resolved).
• Opinion: draft brand GIFs on Flare, finalize on Sunburst.

What shipped in Grok Bot

Phone calls are official. On Wednesday at 1:13 PM ET the official @bot account posted "Grok Bot can now make phone calls." That upgrades earlier user reports. Our story walks the promo: an approval card with the number, a live transcript, Take over and End call, then a recording. Elon Musk followed up Thursday that a Bot "can call your dentist on the phone!" The changelog still tops out at v0.78.1 (Oct 27) with no calls entry, and the promo uses a 555 number. We have not placed a call ourselves.

Before the announcement. Our Tuesday tip covered Dylan (@dkmitc) having his Bot phone Progressive about glass coverage while he picked up his kids. Treat every call as you on the line: say what it may share, what it may not agree to, and read the plan back before dial.

Usage meter. Over the Nov 1 weekend several users said their weekly meter about doubled, then mostly came back Monday (story). A Billing and Usage Dashboard degradation (ps_838fe2a3a154acfa) ran that morning. Export your usage and keep the file if the meter still looks wrong.

No desktop changelog this week. v0.78.1 from Oct 27 was still the newest entry on the changelog Friday afternoon. Separately, Cursor's download server answered our HEAD for macOS arm64 builds labeled 0.79.0, 0.79.1 and 0.79.2 (last-modified Nov 5 UTC) and still serves a 1.0.0 path dated Oct 27. Nothing on the changelog names those builds. Treat them as files on a server, not a shipped release.

How people used it. We published five workflows and two playbooks this week: Muse Gmail draft handoff and a nonprofit support inbox (Nov 2), a daily X teardown (Nov 3), ChatGPT scheduled morning RSS and a T3 Code Notion morning day plan (Nov 6); plus a founders Monday CFO Bot (playbook) and SpaceXSI's shared Data Bot, which @bot said handled "an estimated 70 hours of work that would otherwise have gone to our data scientists" (playbook).

What broke, and what got fixed

• Nov 1. Billing and Usage Dashboard degradation (ps_838fe2a3a154acfa), about 1:38 AM to 5:53 AM ET. Tied to the meter reports above.
• Nov 2. Source control integrations briefly broken (ps_dece7cc03ecef152), about 4:09 to 5:24 AM ET.
• Nov 3. Slow Cloud Agent and Bugbot starts from a GitHub Git operations incident (ps_b88ca87004b1d3b9), about 4:40 to 5:53 PM ET.
• Nov 4 (Wednesday evening). Grok Bot degraded performance (ps_c0cee9e504398968), about 7:34 to 8:49 PM ET Nov 4.
• Friday. Computer use failed for users in a few regions from 6:49 to 7:03 AM ET (ps_183910c201a91593). Grok Bot degraded from 8:57 to 10:15 AM ET (ps_9a160a3b12ec99d1). Colin, from Cursor staff, wrote on the forum thread: "https://status.x.ai stayed up to date during the incident, and was resolved just a moment ago!" Status was all operational when we checked Friday afternoon.
• Still open from last week. Typing "stop" to one Bot can still stop Bots it handed work to. Mohit, from Cursor staff, wrote that it is "a known issue we’re already tracking" and that "the Stop button in that bot’s chat (shown while it’s working) stops only that bot and its own background tasks". Silent-at-limit bots should now say so after you update; deanrie wrote: "Now, instead of silence, it should show a clear message that the limit has been reached."

Cursor

Build Live Apps. The Nov 5 changelog says you can ask a cloud agent to build an app, set up a database, configure auth and deploy it, then share a link. Security checks run before publish; nothing goes live until they pass and you publish. Our story has the steps. Paid plans with cloud agents. We have not published an app ourselves.

Voice Chat with Agents. The Nov 6 changelog adds a realtime voice call with agents in the Agents Window, IDE, web and iOS, on Grok Voice. It is separate from dictation. Our Monday draft covers it for the Nov 9 issue; it is not on the site yet. Until then, read the official entry.

Carry-overs. The Agent SDK remains generally available. Cursor 3.28 on Windows still returns an unexpected error on every agent prompt. Colin wrote: "Just bumping this information that affected users should downgrade to 3.27!" Staff say a fix is identified, with no ship date in that thread.

Beyond Grok Bot

Anthropic. Claude API message threads hit public beta (about 35% lower median time to first token in Claude Code, per Anthropic). FlashProve is an open-source prototype that lets a plain CPU check that a reply came from committed weights. A pass proves less than it sounds. On Thursday Claude announced Personal Finance in beta in the US on Pro, Max and Ultra: "Access is read-only, and Claude can't move, manage, or invest your money" (story). Enterprise got a beta security center. Anthropic also published a scientists usage report.

OpenAI. From December, ChatGPT Plus, Pro and Business chats and Codex share one token-based allowance; Instant chat does not draw from it (story). Images 2.6 adds native GIFs, reusable references and GPT-Image-2.6 Flare and Sunburst. GPT-6 Cyber is for select Daybreak Red orgs with separate model approval. Codex CLI 0.167 opens the agent command center without a daemon. We also profiled Tibo Sottiaux.

Hark. Users said the Nov 1 reset moved their free limit to Dec 1 (story). A user ran three real Mac tasks and timed 5 to 17 minute runs. Dale Thomas says Hark caught hotel and flight mismatches on the free plan (user report).

This week's opinion

Our Friday column, Build a Flare-draft, Sunburst-final GIF brand kit, argues that brand GIF work needs a draft lane and a final lane. Its top pick is a small Images 2.6 runner that drafts on Flare at low complexity, then finalizes looping GIFs on Sunburst behind one approval gate. Second is a Claude Enterprise sign-out checklist that leaves a containment receipt. Third is a Hark usage log only Hark can ship.

What to do now

1. If you try phone calls, write the stop list first: no cancel, no accept, no PIN, no card. Show the number on the approval card, then Call. Read the transcript before you act (story).
2. Export this week's Grok Bot usage and keep the file if the meter jumped on Nov 1 (story).
3. Stop one Bot with its Stop button, not by typing "stop", until staff fix the known cascade.
4. On Windows, stay on Cursor 3.27 if 3.28 breaks agent chat, as Colin advised.
5. Try Build Live Apps on a throwaway internal tool behind team sign-in before you publish anything customer-facing (story).
6. Update Cursor if you want Voice Chat; read the changelog and wait for our Monday write-up if you want our steps.
7. On Claude's US Pro/Max/Ultra, open Personal Finance only if you are fine linking a bank read-only through Plaid (story).

Prompt (ours, untested):

Look back at everything you ran for me from Monday Nov 2 to now. List any routine or task that failed, went silent or stalled, with the day and time, and say what it was meant to do. For each one, tell me whether rerunning it would send anything to anyone, place a phone call, publish an app or spend money. Draft a rerun plan and wait for my OK before you rerun anything.

Why it matters at work

Score 7/10 (official phone calls). Saves: hold music and "call us for details" lookups, about 15 to 60 min a week for errands that stall on a website (our estimate, from user reports and the @bot promo). For: ops, founders, anyone who owns vendor or benefits calls. Setup: not sourced; plan and country limits still unpublished. Connects: phone network via Grok Bot. You stay on the approval card and Take over.

Sources

• Grok Bot changelog, checked Nov 6, 4:20 PM ET; Cursor changelog, checked Nov 6, 4:20 PM ET
• @bot phone calls, Nov 4; Cursor Build Live Apps; Cursor Voice Chat
• Status: summary; incidents ps_838fe2a3a154acfa, ps_183910c201a91593, ps_9a160a3b12ec99d1
• Cursor forum: Colin, Nov 6; mohitjain, stop; deanrie, silent limit; Colin, 3.27
• @claudeai Personal Finance; @OpenAI Images 2.6; Lauren Tan, Slack
• Cursor Workshops, checked Nov 6
• Every linked story carries its own sources.

Today's lead

Cursor agents can now talk with you in a realtime voice call (Cursor · confirmed)

Start a two-way voice call with a Cursor agent in the Agents Window, IDE, web, or iOS, and hear short spoken updates while it codes.

You can now hold a realtime voice call with a Cursor agent. Speak a change out loud, ask what it is doing, and hear short spoken updates while coding keeps running. It is live in the Agents Window, the IDE, on web, and in Cursor for iOS.

What changed

Cursor's changelog is dated Nov 6: "You can now talk with your agents in a realtime voice call."

This is not dictation. Dictation still turns speech into text in the chat input. Voice chat is a two-way call where the agent answers out loud. The docs say the voice agent hands coding work to the coding agent in that chat and keeps talking while the work runs. The changelog says the coding agent "shares short progress notes while it works and a brief answer when it's done". Calls run on Grok Voice, and the changelog adds: "With Privacy Mode on, voice chat runs with zero data retention."

Why it matters

Dictation still leaves you reading the full reply. Voice chat lets you supervise an agent away from the keyboard: ask about a diff, check a run, and usually hear a short spoken update instead of the full written reply.

Why it matters at work

WHY IT MATTERS AT WORK · score 6/10 Saves: hands-free status checks while an agent codes (our estimate: about 15 to 60 min a week for heavy agent users) For: engineering, founders, PMs Setup: update Cursor · connects: none

How to use it

1. Update Cursor and open an agent chat (IDE chats need at least one message first).
2. Clear the input, wait until the agent is idle, then click Start voice chat next to the dictation mic.
3. Allow the microphone if asked. On iOS you can say "Hey Siri, call Cursor" with the phone unlocked (docs).
4. Talk, then hang up when you are done.

Say this out loud (ours, untested):

Walk me through the last diff you made, then add a failing test for the edge case we just discussed and tell me when the run finishes. Keep answers short while you work.

We did not try this hands-on (no Cursor mic session on the box this run).

Good to know

• Signed-in account required. API-key sessions cannot start voice chat; the call fails with "Realtime voice requires sign-in" (docs FAQ).
• The waveform button shows only when idle and the input is empty. It is hidden for subagents, Project workers, archived chats, Claude Code chats, and teammate-owned chats.
• About 10 minutes of silence with no work hangs up. There is a daily call limit.
• @cursorlog at 2:15 PM ET: "Cursor Voice Chat has been released." We found no @cursor_ai post about it as of 2:26 PM ET.

What to watch

A staff post on @cursor_ai, the exact daily limit, and reports of a missing Start voice chat button after update.

Last checked: Nov 9, 2026, 6:29 to 6:31 AM ET. What changed: no change since Nov 6, 2026, 2:24 to 2:28 PM ET. Reopened the Cursor Voice Chat changelog, docs, forum.cursor.com/t/176354 and forum.cursor.com/t/177703; changelog still dated Nov 6 with the same quotes; docs FAQ and limits unchanged; 176354 still shows Grok Voice calls in IDE 3.27.22 with a Voice call card (Colin, Oct 27); 177703 is a Nov 8 first-impressions thread with no staff reply.

Sources

• Cursor changelog: Voice Chat with Agents
• Cursor docs: Voice chat
• Cursor Changelog (@cursorlog) on X

What to do:

1. Update Cursor.
2. Open an agent chat. In the IDE, send at least one message first.
3. Clear the chat input and wait until the agent is idle so the waveform button replaces Send.
4. Click Start voice chat (next to the microphone used for dictation) and allow the mic if asked.
5. Talk. Ask the agent to make a change, check a diff, or report on a run. Hang up when you are done.

Prompt (ours):

Walk me through the last diff you made, then add a failing test for the edge case we just discussed and tell me when the run finishes. Keep answers short while you work. (ours, untested)

What's new

ChatGPT dots can now use sign-in codes from your inbox (OpenAI · confirmed)

Your dot can fill an emailed sign-in code from your connected inbox once you allow it, but the always option covers any website.

Your ChatGPT dot can now autofill emailed sign-in codes, per OpenAI's James Sun, who posted Saturday at 2:17 PM ET: "A little upgrade: your dot can now use sign-in codes sent to your connected inbox to log in". @ChatGPT reposted it.

What changed

OpenAI's Learn page documents a private form: "Enter your credentials there and complete any verification the website requires, such as a verification code." Now, per Sun: "When prompted, you can let your dot use the code once or always allow it to fill in." In his demo video, the approval prompt reads "Use the verification code sent to your connected email to sign in to resy.com. ChatGPT cannot see the code." Product wording, untested by us.

Why it matters

Our read: autofill is an alternative to the documented private-form and takeover routes, not guaranteed login completion.

How to use it

1. Connect your email. On mobile, per OpenAI's Help Center, "open your dot’s profile and go to Customize > Plugins. Select a supported app and review its permissions before connecting it."
2. Ask your dot to sign in to the site with your email.
3. At the prompt, pick Allow once, Always allow for any website, or Cancel (demo labels).

Ours, untested:

Sign in to [site] with my email address. If the site emails a sign-in code, ask me first and use it once only. Then tell me what you see on my account page.

Good to know

• Asked about two-factor codes from a password manager, Sun replied Sunday: "Only works for email OTP right now".
• The demo option says "Always allow for any website", not every action. Our read: prefer Allow once.
• Neither Sun nor the two docs above list plans or mail apps for autofill; general dot eligibility is in Help.
• Per Learn, existing app permissions apply: inbox access isn't code-only, and read permission doesn't authorize sending. Per Help, "Disconnecting an app does not delete information your dot has already obtained from it."

What to watch

Sun said Sunday: "we are working on expanding the ability for more types of automatic agent auth with other formats like magic links, passkeys, etc". Watch the Learn page's "Sign in to a website" section.

Last checked: Nov 9, 2026, 5:59 to 6:06 AM ET. What changed: added Sun's Sunday OTP-only reply and Help's disconnect note.

Sources

• James Sun (OpenAI) on X: dots can use sign-in codes from your inbox (Nov 7, 2026)
• @ChatGPT on X: repost of James Sun (Nov 7, 2026)
• James Sun (OpenAI) on X: only works for email OTP right now (Nov 8, 2026)
• samir (OpenAI) on X: just rolled this out (Nov 7, 2026)
• ChatGPT Learn: Connect computers and apps to your dot
• OpenAI Help Center: Getting started with your dot

What to do:

1. Connect your email: on mobile, open your dot's profile, go to Customize > Plugins, select a supported app and review its permissions, per OpenAI's Help Center.
2. Ask your dot to sign in to a website with your email.
3. When prompted, let your dot use the code once or always allow it, per OpenAI's James Sun.

How people are using it

How users are moving Grok bot jobs into Hark (Hark · reported)

One user moved Grok bots into Hark with handover notes and Scheduled Tasks; our untested plan maps jobs to Projects, Specialties and Scheduled Tasks.

What you can do now: one user says they moved their Grok bots into Hark on Saturday. On Sunday they clarified: ask each Grok bot for a handover note, give those notes to Hark, and turn most into Scheduled Tasks. Our untested suggestion: paste old bot instructions into Hark, ask for a Project / Specialty / Scheduled Task plan, and approve one by one. Three users describe other setups.

What changed / the reports

On Saturday at 6:49 AM ET, Власенко (@vlasenkoyakov) wrote: "I just moved my whole crew of Grok bots to @hark_labs, and the best part: Hark did the move itself." The same post says: "It read my old bots, rebuilt each one, kept every rule I'd set, and asked only where it truly needed me." (user report)

Jobs named include "My weekly work report, drafted from short notes I send during the week, in my own style" and "My tax and benefit deadlines, watched quietly so nothing slips", plus weekly mail and drive cleanup and a Friday music list. (user report)

Brett Adcock, Hark's founder, reposted it at 11:21 AM ET Saturday (founder repost). The poster had 34 followers Monday (~12,860 views). A founder repost is not independent verification, and the Saturday post does not show every rule carried over.

Other reports we found:

• Alex Findley (@AlexFindleyAI) wrote at 2:25 PM ET Saturday: "I was running 25 bots. Today I cut it to 18 and moved the heavy lifting elsewhere." He says Hark "writes the briefs, checks every pull request, and keeps my money, goals and diary up to date" and closes: "The bots still have a job, just fewer of them doing less." (user report) Earlier at 1:16 PM ET he replied: "Completely done away with grok bot for Hark now" (user report). The two posts do not match, and he has not explained the difference in what we checked through Monday morning.
• Amine (@AmineTX) replied at 11:47 AM ET Saturday: "I use Hark as a back up when I ran out of bot credits each of my bots will give hark a hand off file." (user report)
• Daksh (@ddakshh) wrote at 4:05 PM ET Friday: "Today I set up Hark projects as department bots, with a “Chief of Staff” coordinating them all." (user report)

Cory Harelson (@CHFIG132) asked at 7:27 PM ET Saturday: "How did you set this up? Is each bot a project?" (user report). On Sunday at 11:05 AM ET, Власенко replied: "Mostly not projects! Each bot became a scheduled task: its own instructions, schedule and memory notes, e.g. a music scout every Monday, a weekly log on Friday. Bigger ongoing stuff (work, taxes) got its own project. Tip: paste your old bot's instructions into Hark and say 'run this every Friday'. That's basically it 🙂" (user report). At 11:03 AM ET they wrote: "No magic, just paperwork 😄 I asked each Grok bot to write its own handover note (what it does, its rules, its routines) and gave those to Hark. Like an exit interview for bots. Hark took it from there." (user report). We did not run a migration, so the plan below is ours and untested.

Why it matters

Adcock, Hark's founder, posted at 10:16 PM ET Sunday, Nov 1: "For those of you struggling with the Grok Bot usage limits, Hark’s tiers are incredibly generous. Give it a try" (founder post). This weekend's posts name recurring reports, deadline watching, cleanup and backup work.

How it maps to Hark's docs

Hark's Using Hark page names three places a moved bot could live:

• Projects: "dedicated threads for longer term work. They focus a single topic, and related files, in one place that you can visit again and again over time."
• Specialties: "workflows that Hark memorizes to repeat a multi-step task you reliably over and over again," set up automatically "or you can ask to create one."
• Scheduled Tasks: "You can create them just by asking Hark."

No import section is on that page. Treat Saturday's "Hark did the move itself" as one user's account, clarified Sunday as handover notes plus Scheduled Tasks and Projects. The mapping below overlaps that tip; we have not tested it.

How to use it

1. Copy each old bot's instructions and schedule into one note from that tool's settings.
2. Paste the note into Hark and ask for a plan mapping each bot to a Project, a Specialty, or a Scheduled Task.
3. Read the plan. Approve one bot at a time, starting with a low-stakes job.
4. Connect only the accounts that bot needs, and decide whether Hark may change anything without asking.
5. Check your usage after the first few days. Credits can still run out.
6. Keep the old bot running until the Hark version finishes one full cycle, such as a weekly report.

Prompt (ours, untested):

I want to move some of my old assistant bots into Hark. Below I pasted each bot's name, its instructions, and when it runs. For each one, tell me whether it should be a Project, a Specialty, or a Scheduled Task in Hark, and list any rule you can't carry over. Do not connect new accounts, send anything, or spend money. Show me the plan and wait for my OK before you create anything. Tell me which rules you are not sure you can carry over exactly.

Good to know

• Projects are not separate bots. Hark's docs call Projects "basically just Chat, but organized into individual topical threads" and say Hark "learns from your conversations, and it applies that context to everything you ask it." That is Hark's description; we did not test it. Where Memory is on, assume what one Project learns can reach the others. The privacy policy says you can "turn Memory off there at any time" and that "Memory is not available in the European Economic Area" (UK and Switzerland too). A prompt is not an enforced boundary, so do not move material two bots must keep apart. FoxKeyn (@Foxkey_) replied in Chinese at 10:25 PM ET Saturday that Hark memory is shared, unlike separate Grok bots (user report, our translation).
• Usage can run out fast. MJ (@MJTraderPro) replied at 4:24 PM ET Saturday: "You’ll rip thru credits even faster with hark. I am maxing out every 2-3 days on the highest plan" (user report). Власенко replied Sunday they had "used about 3% of my monthly pool", with "Long browser sessions" burning credits while scheduled tasks stayed light (user report). Pricing lists Pro² at $20 and Pro³ at $100 a month. Auto top-up is optional; if chosen, Hark refills "with an amount you decide, up to a ceiling you predetermine." Credits can still run out. Free plan: see our Dec 1 reset story.
• Scope the access. Findley says "Hark got full access to my app" and "Every change it makes can be undone". That is his app setup, not a Hark control. Give least access each job needs.

What to watch

Whether Findley clarifies his two posts, whether Hark documents imports or per-Project memory, and whether more users show the handover-note path or rule-for-rule preservation.

Sources

• Власенко (@vlasenkoyakov) on X: moved my whole crew of Grok bots to Hark (user report)
• Brett Adcock (@adcock_brett) on X: repost of the Grok bots move (founder repost)
• Alex Findley (@AlexFindleyAI) on X: cut 25 bots to 18, Hark does the heavy lifting (user report)
• Amine (@AmineTX) on X: bots hand Hark a handoff file when credits run out (user report)
• Daksh (@ddakshh) on X: Hark projects as department bots with a Chief of Staff (user report)
• FoxKeyn (@Foxkey_) on X: Hark memory is shared, unlike separate Grok bots (user report)
• MJ (@MJTraderPro) on X: maxing out every 2-3 days on the highest plan (user report)
• Cory Harelson (@CHFIG132) on X: How did you set this up? (user report)
• Власенко (@vlasenkoyakov) on X: handover notes from each Grok bot (user report)
• Власенко (@vlasenkoyakov) on X: mostly Scheduled Tasks, not projects (user report)
• Власенко (@vlasenkoyakov) on X: used about 3% of my monthly pool (user report)
• Brett Adcock on X: Hark’s tiers are incredibly generous (founder post, Nov 1)
• Hark: Using Hark Pro (Projects, Specialties, Scheduled Tasks, Memory, auto top-up)
• Hark: Pricing (free full features, Pro² $20, Pro³ $100)
• Hark: Privacy Policy, last updated Nov 5 (Memory section)
• Alex Findley (@AlexFindleyAI) on X: Completely done away with grok bot for Hark now (user report)

Last checked: Nov 9, 2026, 6:12 to 6:18 AM ET. What changed: Monday carry refresh; added Власенко's Sunday handover-note and Scheduled Tasks clarifications; Findley conflict still unresolved in checked posts; Using Hark, Pricing, Privacy (still Nov 5), and status.hark.com unchanged; no Adcock/@hark_labs ship yet for the teased Monday update.

Make Cursor refuse agent shell commands outside the sandbox (Cursor · confirmed)

Block unsandboxed shell commands with Colin's hook; folder read limits still come from Cursor's sandbox settings.

Refuse unsandboxed agent shell commands with Colin's hook. It checks sandbox, not cwd or paths.

What changed

Friday's settings required approval outside the sandbox; Auto-Review could approve automatically. Saturday's hook refuses shell requests instead. Read Access = Workspace supplies the sandbox's folder read boundary.

Why it matters

For repo work, the hook prevents Auto-Review from approving unsandboxed shell execution. It does not guard file edits, reads or MCP tools.

WHY IT MATTERS AT WORK · score 2/10 Saves: no measured time saving. For: engineering, security. Setup: unmeasured; Python 3 required · connects: none.

How to use it

1. Follow post 5: Settings > Agents > Execution and Approvals, Allowlist (with Sandbox), empty allowlist, Read Access = Workspace.
2. Create ~/.cursor/hooks/; save the full Python script from post 12 as stay-in-sandbox.py there.
3. Create ~/.cursor/hooks.json with Colin's config below. Existing file? Append the entry to beforeShellExecution, don't replace other hooks. Save; no restart.

{
  "version": 1,
  "hooks": {
    "beforeShellExecution": [
      { "command": "python3 hooks/stay-in-sandbox.py", "failClosed": true }
    ]
  }
}

Our prompt (untested):

Run echo hook-test with the sandbox turned off. If Cursor refuses, quote the exact message.

Expected hook message: "Blocked: this command would run outside the sandbox."

Good to know

Shell only. Wider-network requests and allowlisted commands are refused too. Sandbox off blocks everything; failClosed blocks on script failure, including missing Python. User hooks don't run in cloud agents. No hands-on: no Cursor IDE available on our computer.

What to watch

Watch 177516 for a staff reply announcing a built-in refusal setting.

Last checked: Nov 9, 2026, 6:32 AM ET.

What changed: no change since Sat Nov 7, 7:05 AM ET check (forum 177516 still 4 posts through Colin #12; hooks docs and both changelogs unchanged; status.x.ai all systems operational).

Sources

• Colin's hook and full script
• Colin's settings and allowlist limits
• Cursor hooks reference

Workflows and Setup of the day

Grok Bot for sales ops: three shared Bots, call to order form (Grok Bot · reported)

SpaceXSI's sales team split its deal cycle across three shared Bots that work with each rep's own permissions and leave judgment calls to people.

You can give a sales team a few shared Grok Bots, one per job in the deal cycle, so reps spend less time updating Salesforce after calls and chasing quotes through approvals. On Oct 27, SpaceXSI's sales team described running three: Account Intel Bot for prep and follow-up, Quote Bot for proposals and order forms, and Deal Bot for its Deal Desk (SpaceXSI). It fits sales ops and deal desk leads with a CRM and a written discount policy. It's the company's own account.

Not tested: we have no Salesforce org, CPQ tool or Cursor team to run it against.

Who runs it

The post was bylined Iris Wang, Joe Robertson, Micheal Sihavong and Palmi Angelova. Wang's X bio reads "building deal desk with @bot @SpaceXSI. prev @cursor_ai, @plaid, @mckinsey".

Disclosure: SpaceXSI makes Grok Bot, so this is the vendor describing its own product in use. The page ends with "Try Grok Bot today" and a Contact sales button. The official @bot account promoted it on Oct 27 at 1:17 PM ET, and Wang called Deal Bot "my baby". The post says Team Bots; the help pages say Shared Bots.

How the sales team set it up

From the Oct 27 post, in SpaceXSI's words:

1. One Bot per existing role. "We began creating Team Bots, each specializing in a role that already exists in our Sales organization. The people who know each job best help build and maintain the Bot".
2. Let the logs pick the first job. "Our GTM engineering team started by having Grok Bot analyze Salesforce activity logs to identify the most valuable work to automate." That showed "reps were spending a lot of time updating opportunities after customer calls."
3. Limits first. "They act with each user's permissions, with additional restrictions on sensitive actions. We've built rate limits to prevent unintended changes at scale and reserved certain actions for specialist Team Bots." The post's example: "A sales rep, for example, cannot use Quote Bot to approve their own discount."
4. People review the fuzzy parts. "For now, we require human review whenever we can't clearly define the criteria for a decision, or when a Team Bot prepares customer-facing communications."

The deal cycle, Bot by Bot

• Account Intel Bot. "Every rep has a private conversation with the same Team Bot". It "can prepare a briefing based on account history before a customer meeting. Afterward, it can read the Gong transcript, identify and make the necessary Salesforce updates, and draft a follow-up email for the rep to review." And: "Reps can schedule these workflows to run automatically".
• Quote Bot. It can "prepare several proposals with different products and discounts", puts them in Salesforce and shows them side by side in the app. After approval it "can generate the order form in Ironclad, download the executable contract, and draft an email for the rep to send to the customer." The guardrail: "Before changing anything in Salesforce, it shows the rep a preview and asks for confirmation."
• Deal Bot. It "checks requests against our approved policies and contract language," can "monitor and triage Slack requests submitted by reps," and checks whether Legal has drafted similar terms. It holds "permissions that sales reps don't", but for judgment calls "it prepares a recommendation and asks a Deal Desk specialist to approve it before proceeding." Each approval it processes "is stamped in the system with the decision, the rationale, and the DRI whom Deal Bot was supporting".

The @bot post diagram showed the handoffs: Account Intel "Requests quotes" from Quote Bot, which "Escalates exceptions" to Deal Bot, which "Returns approved terms".

Still running? The post described it in the present tense and named two more Team Bots, "Competitive Intel Bot" and "Sales Training Bot". On Oct 27 Wang wrote "more bot teammates coming soon!" We found no later update.

Why it matters at work

Our read: time goes to CRM updates after calls, quotes in a queue, and discount requests lost in Slack. SpaceXSI split those jobs along its org chart, so the Bot that drafts a discount can't approve it. The only figure is Wang's: "Deal Bot has already helped my team save hundreds of hours this month." She gave no counting method, and the post gives no accuracy or override rates.

Set it up

These steps are ours, built from the post and Cursor's help pages. The post doesn't share its skills, prompts, routine settings or how each tool is connected.

1. Find the first job in your own logs. See which CRM update reps make most after calls.
2. Create one Shared Bot for that job only. Shared Bots "are for members of a Cursor team who have Grok Bot" (Shared Bots help).
3. Write the rules as skills. Which fields change after a call, when a stage moves, and what your discount policy allows.
4. Make the process owner a Manager. Managers can change "its instructions, skills, memory, and apps" (Shared Bots help).
5. Keep approval with the approver. Our read: give the Deal Desk Bot its own access, and keep reps' Bots unable to approve.
6. Let reps set their own routine. Test it on a sandbox account first, since "A test does real work" (Routines help).

Starter prompt (ours, untested)

You are [Account Intel Bot] for our sales team. After each customer call on my calendar, read the [Gong] transcript and the account's history in [Salesforce]. Draft the opportunity updates our rules call for (next step, stage, close date, risks), each with the line from the call that supports it, and draft a follow-up email to the customer. Show me the updates as a preview here. Don't change anything in Salesforce until I reply OK, and never send the email, change a price or discount, or submit a quote for approval. Those stay with me.

The transcript, account history, Salesforce updates and follow-up draft are from the post, and the preview before any Salesforce change is Quote Bot's rule there. The calendar trigger, the field list, the supporting line for each update and the never-send and no-discount rules are ours.

Good to know

• Whose login does the work? "A shared Bot uses the apps and secrets its owner added. Those work the same for everyone who chats with it." But "A connected account acts as the person the Bot is answering" (Shared Bots help). The post doesn't say how SpaceXSI wired Salesforce. Our read: to match the post's "each user's permissions", don't put an admin Salesforce login on a Bot every rep can use.
• Slack is different. "In Slack channels, threads, and group chats, a shared Bot uses only its owner's apps, and asks before it uses your accounts." In a thread, "Only the person who asked can answer" its approval card (Shared Bots help).
• Routine approvals expire. When a routine raises an approval card, "it expires after 30 minutes" (Approvals help). Our read: a rep in back-to-back calls can miss one.
• Pin rules with Ask first. "An action that matches one of your Ask first rules always gets a card", and "If two rules conflict, Ask first wins" (Approvals help).
• The rep pays. "A shared Bot reply uses the Grok Bot usage of the person it's answering" (Shared Bots help), and each routine run "spends usage" (Routines help).
• Deleting it takes the reps' routines. Only the owner can delete a shared Bot, and "Teammates lose it, along with their chats and routines on it. It can't be undone" (Shared Bots help).

What to watch

• Numbers. No accuracy or override figures yet. Signal: a SpaceXSI follow-up with how often reps or the Deal Desk changed a Bot's work.
• More autonomy. "we expect to give our Team Bots more responsibility." Signal: a post saying a Bot now sends customer email or approves terms without review.
• The next Bots. Signal: a write-up of Competitive Intel Bot or Sales Training Bot.

Related

• Grok Bot for sales: a playbook for prospecting, sequences, and pipeline
• Grok Bot for sales: meeting to proposal, signature, invoice and chase
• Team Bots enter public beta

Last checked: Nov 9, 2026, 6:29 to 6:31 AM ET. What changed: Shared Bots help now reads "A connected account acts as the person the Bot is answering" (dropped "always" since Nov 5 check). SpaceXSI post, Iris and @bot posts, Routines and Approvals help otherwise unchanged.

Sources

• SpaceXSI: Your next sales teammate should be a Grok Bot (Oct 27)
• @bot on X: Our sales operations team built Shared Bots (Oct 27)
• @irislovesdeals on X: how we're using Grok Bot on deals (Oct 27)
• Cursor Help Center: Shared Bots
• Cursor Help Center: Routines
• Cursor Help Center: Approvals and Auto-review

The digest

Grok Bot and Cursor

• Cursor on Windows: agent won't start on 3.28 and 3.29; staff say roll back to 3.27. A user reported every chat failing with "The agent couldn’t start because Cursor’s agent process didn’t respond." Colin replied at 4:17 AM ET Monday: "this is a known issue on Windows with recent Cursor versions (3.28 and 3.29)." He added: "For now, installing 3.27 from Cursor · Download gets the agent working again." To stop it updating back, he says to run "Preferences: Open User Settings (JSON)", add "update.mode": "manual", save and restart Cursor; managed installs may have that setting locked. Staff member tarun later asked the same user for logs. Good to know: our Nov 7 digest pointed proxy users to 3.28; Windows users on 3.28 may hit this. We found no 3.28 or 3.29 entry on the Cursor changelog page. (confirmed, staff; how many Windows users are hit is not stated)

• Grok Bot can't publish to Cursor Apps on personal plans. A user asked why cloud agents started by Grok Bot see no Apps publish controls. Colin said at 4:29 AM ET Monday: "On a personal plan (no team), Apps publishing is only available to cloud agents started from Cursor itself. Agents that Grok Bot starts don’t get the publish tools". His fix: "start the change (or a follow-up on the app) from a chat in Cursor, and use the Preview/Publish tray there." He did not say whether team plans differ. (confirmed, staff)

• Grok Bot shrinks MCP images to 1024 px; ask the tool for a link instead. Colin wrote at 4:39 AM ET Monday: "Right now Grok Bot resizes every image an MCP tool returns so its longest side is at most 1024 px, before the bot sees it." The saved copy is the resized one. His workaround: "have the tool also return the full-resolution JPEG as a download link (or a file path the bot can reach) in its text output." The bot can then download the original with a shell command. He passed it to the team. (confirmed, staff)

• Cursor cloud agent setup scripts pasted on Windows break on line endings. A user's install script failed with $'\r': command not found. Colin explained at 4:27 AM ET Monday: "This is a known issue on Windows: the web editor saves line breaks as CRLF, and the extra \r gets passed straight to bash." His workaround: commit the steps to a repo file such as .cursor/install.sh with LF endings (*.sh text eol=lf in .gitattributes), set the Install field to just bash .cursor/install.sh, do the same for Start, save and start a new build. The reporter said the workaround is good. (confirmed, staff)

• Grok Bot on Chromebooks: chat works, local commands don't. Colin said at 4:22 AM ET Monday the helper that runs commands on your computer "needs a newer C++ runtime than Debian 12 provides, and Debian 12 is the default for the ChromeOS Linux container." He added: "Installing a newer Grok Bot version doesn’t fix this yet." To check, run tail -n 30 ~/.grokbot/local-exec-daemon.log; head -n 3 /etc/os-release and look for GLIBCXX; the reporter confirmed a GLIBCXX_3.4.31 error on Debian 12. For now he says local commands work from another computer on the same account, including Linux on Debian 13 or Ubuntu 24.04 or newer. (confirmed, staff plus one user log)

• Cursor Projects ignore the Default Environment setting. Colin said at 4:10 AM ET Monday: "Default Environment only applies to New Chat outside a Project." To run a Project agent on your machine, click New Agent in the Project, open the picker next to "New Agent in" above the composer, choose your machine under Remote Machines, then send. The reporter confirmed it runs after picking the machine and asked for a per-Project always-local option; no staff answer on that yet. (confirmed, staff)

• Cursor agent HTTP/2 traffic now goes through your proxy. Colin from Cursor staff wrote at 6:16 AM ET Saturday: "Agent HTTP/2 traffic now goes through the configured proxy (both http.proxy and HTTPS_PROXY) as a CONNECT tunnel." He said it "shipped in a recent Cursor update." On 3.28, which he called the latest version, remove "cursor.general.disableHttp2": true from your settings, then fully quit Cursor (Cmd+Q on a Mac) and reopen it. That replaces the Oct 29 staff workaround of turning HTTP/2 off on proxy-only networks. If you still see direct connections to api5, Colin asks you to say so on the thread. Windows users: staff now say the agent may not start on 3.28 and 3.29; use 3.27 until that is fixed (see today's Windows rollback item). (confirmed, staff)

• Grok Bot browser loopback block is intentional. Colin from Cursor staff confirmed at 5:32 AM ET Saturday that the loopback block is intentional for bot computers rebuilt on the current image. Without it, any page open in the bot's browser could connect to services on that computer. The browser can now use only its built-in proxy on loopback. He said the shell is unaffected; that describes current behavior, not a recommendation to bypass the browser restriction. Private/Tailscale-address egress is a separate question that his reply does not address. (confirmed, staff)

• Grok Bot cloud routing: staff explanation confirmed, user refusals reported. Cursor staff member mohitjain said at 4:12 AM ET Friday that repo edits default to cloud agents. Local work needs an explicit ask each time, has no standing setting and excludes scheduled runs. Some users say the explicit ask still fails: pellinore wrote at 5:44 PM ET Friday, "Even when I have explicitly asked it to do so." d3n4c3 reported continued refusals at 12:03 AM ET Saturday. At least two users, d3n4c3 and aridez, reported canceling. (staff explanation confirmed; refusals and cancellations are user reports)

• Update, Nov 8: staff steps for a suspected startup-file hang. Follow-up to our Nov 6 item. For one user's offline Bot computer, Colin wrote at 6:44 AM ET Sunday: "Reset and Update don’t clear it because they bring back the same startup files." He suspects a startup-shell hang, unverified for this machine. His steps: tell one bot "Don’t run any commands." Test a small copy to and from its computer; if that hangs, stop and reply in the thread. Otherwise copy the /etc/skel defaults over /home/box/.bashrc and /home/box/.profile, wait a minute, run Update Bot's Computer once, then ask a bot to run echo ok. Custom lines in those files are lost. The original reporter, John_Porter, then replied at 8:49 AM ET Sunday, "That worked. Thank you!" (staff recommendation confirmed; reporter says steps worked; cause still suspected)

• Cursor: GLM 5.3 Flash fails on chats that include an image. Colin said at 6:47 AM ET Sunday it's "a known issue the team is aware of": since Saturday, GLM 5.3 Flash returns the provider error "whenever the conversation includes an image." Text-only requests still work. An image anywhere in the chat's history is sent again with each message, so start a new chat and pick another model, or stay on GLM 5.3 Flash without images. (confirmed, staff)

• Cursor Cloud Agents: use a marketplace skill as a mode by copying it into the repo. On Cloud, Use as Mode "currently only pins skills that live in the repo the agent is working on," Colin said at 6:37 AM ET Sunday, calling it a known issue. Workarounds from that post: put the skill's SKILL.md in the repo (his example: .cursor/skills/poteto-mode/SKILL.md), commit, push and start a new Cloud run on that branch; or start each message with the skill's slash command and press Enter (not Option+Enter) to attach it to that one message. Local agents aren't affected. (confirmed, staff)

• Cursor: Allowlist only works on personal accounts, but the dashboard hides it. Colin said at 6:41 AM ET Sunday that environment-level settings, including Allowlist only, work on personal accounts without an upgrade, though the new dashboard doesn't show the control yet. For a repo-managed environment, he says to set "egressMode": "network_settings_only" and an egressAllowlist in that repo's .cursor/environment.json, then trigger a new build; only that environment changes. The allowlist isn't a full block: Cursor's Secrets & Network docs say "Even in Allowlist only mode, a small set of domains remain accessible so Cloud Agents can function. These include Cursor's own services and source control management (SCM) providers." (confirmed, staff and docs)

• Cursor CLI on Windows: one user measured about 7 seconds of overhead per hook. On Windows 11, PowerShell 7.6.9 and CLI 2026.11.05-86047df, a user reported the hook wrapper alone took 6.4 to 7.2 seconds, about 6 of them in a state dump that writes one line at a time. Colin confirmed the mechanism at 6:43 AM ET Sunday: "So every hook pays for the full state dump, even though a hook doesn’t need any of that state." He passed it to the team and wrote: "In the meantime, the Add-Content override you tested is a reasonable stopgap." He said to keep it scoped to those processes. (mechanism confirmed, staff; timings one user's report)

• Update, Nov 8: private gateway bridges can't send with computer credentials. Colin said at 6:38 AM ET Sunday the server now refuses sends made with the computer's own credentials, after a security change. He added: "Messages you send from the Grok Bot app are still going through." GitHub users reported refused sends from Nov 5. Colin recommends saving incoming messages where the Bot can read them, then calling a routine webhook. He says the webhook doesn't pass the body; the help center and Aug 25 release notes say the Bot receives it. The reporter wrote Monday, "The webhook carries no message." Neither explains why. We tested neither the bridge nor which holds; our Oct 8 tip needs that caveat. (staff and docs confirmed as statements; body behavior unresolved)

Known issues:

• Watch: Bots refusing browser tasks and X posting since Nov 6 (reported, no staff reply). Users in one thread say Bots stopped browser work on several sites, including Slack, GoHighLevel and Meta Ads, citing an "access restriction." In another, users say their Bot won't post to X through the browser and points them to the X connector, which per the original poster "can’t create posts with images." No staff reply in either thread as of 6:34 AM ET Monday. (reported, several users; cause unknown)

Anthropic

• Claude Code 2.1.319 is now npm latest. Claude Code 2.1.319 is now npm latest and next, published Nov 5 at 7:58 PM ET on the npm registry (GitHub release Nov 6 at 12:04 PM ET). The changelog says MCP tasks in the terminal move to the background (set CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND to wait instead), and permissions.denyInsteadOfPrompting denies a tool call that would wait on a permission prompt and lets the turn continue. It also fixes a sed command that could skip the permission check in accept-edits and auto mode, a Bash bypass through a crafted environment prefix, and a sandbox escape when creating a worktree. Gateway policies with no desktopHome that rely on coworkTabEnabled: false or chatSessionRetentionDays get a warning ahead of Claude Desktop's undescribed November 10 change. (confirmed)

• Project EVE: profits in a replayed market; instructed sabotage slips past. On Nov 6, Anthropic's Frontier Red Team published Project EVE: Claude plays MMORPGs. In a replayed EVE Online market, not live trading, it says Opus 4.7, Opus 4.8, and Fable 5 were profitable in all nine two-week windows, growing capital by about 16%, 22%, and 43% on average. In live EVE Frontier, four-agent teams coordinated, but teammates almost never caught an agent Anthropic instructed to sabotage them. Limited-access monitors never exceeded 50% confidence. Anthropic says capital-managing agents need explicit risk limits, and that agent teams will probably need audit trails and separation of duties. The tests show instructed sabotage, not an inclination to sabotage, and the replay cannot react like a live market. (confirmed)

• Cloud sessions: ClaudeDevs says env setup got easier. On Nov 6 at 2:58 PM ET, @ClaudeDevs said it is easier to get a repo running in Claude Code cloud sessions on Pro, Max, and Ultra: "Claude figures out what dependencies, network access, and API keys it needs, then saves the environment for next time. API keys are saved securely, so Claude never sees them." The cloud environments docs also list Team plans and Enterprise users with premium seats or Chat + Claude Code seats. They say you can store network secrets that sessions use without seeing them, run a setup script, and use environment caching. The docs do not separately describe Claude auto-detecting those needs. (confirmed)

• Agent memory: dreaming, and a 200-line MEMORY.md cap. On Nov 6, Anthropic's Engineering team published The components of agent memory, and @AnthropicAI posted it at 4:16 PM ET. It describes dreaming as offline reorganization that can correct mistakes, create notes, build new memory representations, or shape what later sessions load. Of Claude Code's auto memory, the post says: "It loads only the first 200 lines or 25 KB of MEMORY.md." It also says: "Injected memories use up valuable tokens, and an incorrect note can be carried into every session." The post says Claude Code documentation warns that longer instruction files reduce adherence. (confirmed)

OpenAI

• Pro plans have no 5-hour limit; staff list 10X, 20X and 50X multiples. Early Sunday at 12:09 AM ET, Tibo Sottiaux (Codex and ChatGPT at OpenAI, per his X bio) posted: "Saw some confusions, but all Pro plans have no 5-hour limits, and give the following usage compared to Plus." His list: Pro 100 "10X", Pro 200 "20X", Pro 500 "50X, access to ultrafast". OpenAI's usage help article backs the first part: "Pro 100, Pro 200 and Pro 500 currently have no five-hour usage limit in Work and Codex. Your plan still has an included usage allowance." Neither that article nor About ChatGPT Pro tiers prints the multiples. (multiples: staff claim)

• Dots: staff say wake/sleep and adherence fixes are in progress. Saturday at 2:30 PM ET, James Sun (browser use for Codex and ChatGPT at OpenAI, per his X bio) told a user who said their dot had been offline for almost 3 days: "We are fixing some wake/sleep issues right now". Monday at 3:11 AM ET, samir (Dots and ChatGPT at OpenAI, per his X bio) replied to a user who wrote, "My dot notes everything and never actually corrects the action or line of thinking": "we are actively working on improved adherence, memory and behavior". Neither says a fix has shipped. Per OpenAI's Help Center: "To report a bug, share product feedback, or request a feature, ask your dot in its main conversation." (staff-confirmed, in progress)

• Sites Remix rolls out to Business and Enterprise. Friday at 5:08 PM ET, @ChatGPT posted: "With Remix, make your own version of a site by telling ChatGPT what you'd like to change. Then keep editing your copy in chat." The announcement names Business and Enterprise only: "Now rolling out to ChatGPT Business and Enterprise accounts." That differs from co-editing, where per OpenAI's Sites help page "An editor can update and save the shared Site." Neither Help page we checked (that one and Business release notes) documents Remix yet. (confirmed, rolling out)

• Digiday: OpenAI is testing a Business Hub mention-count dashboard (unverified). Digiday reported Monday: "OpenAI is testing an analytics dashboard that counts how often a business shows up in ChatGPT’s answers and how many clicks those appearances send to its site." Per Digiday, "The tool sits inside a wider Business Hub that hasn’t launched yet," getting in requires proof of authority to represent the business, and prompts and competitor rankings stay hidden. Digiday cites two sources and a November 6 LinkedIn screenshot from Juozas Kaziukėnas, founder of Marketplace Pulse. "OpenAI declined to comment for this story." We found no OpenAI announcement. (reported, unverified)

Known issues:

• Some U.S. business Mastercards declined; image errors resolved. OpenAI's status incident on payment declines for some U.S. business Mastercards (numbers beginning 544015, 546400, 543502, 543799 or 512998, and some 513989) opened Friday at 1:38 PM ET and has said since Friday 8:29 PM ET: "We’ve applied a mitigation and are monitoring the affected services to confirm recovery." It warns: "If your prepaid balance runs out, your API requests may fail." Separately, elevated image generation errors ran Sunday from 10:11 AM to 12:00 PM ET, now resolved. (confirmed)

Hark

• Adcock says four Hark updates land this week, first Monday morning. Brett Adcock, Hark's founder, posted at 11:20 PM ET Sunday: "Hark has 4 important product updates coming this week, first one tomorrow morning" (founder claim). He named no features. The post had about 36,000 views and 89 replies when we checked at 6:03 AM ET Monday. We found nothing new on hark.com, the app stores or @hark_labs yet, so treat the updates as announced, not shipped. We will check again.

• iOS 2.8.1 now shows 2,800 ratings; Android is on 1.0.0-81. Apple's lookup at 6:02 AM ET Monday shows version 2.8.1 (released Fri Nov 6, 11:58 AM ET), notes "Refinements to Hark", 2,800 ratings, average 4.72. That is up from iOS 2.8.0 with 2,623 ratings in the Nov 6 issue. Google Play's listing, checked at 6:16 AM ET Monday, shows package version 1.0.0-81, "Updated on Nov 8, 2026", about 4.52K reviews. The Mac DMG, checked at 6:16 AM ET Monday by HTTP Last-Modified, ETag and Content-Length (not an install), is unchanged from that Nov 6 line: Last-Modified 1:34 PM ET Thu Nov 5, size 20,465,674 bytes. (store listings)

• Gizmodo names Hark Pro Best AI Software of 2026. Gizmodo's Best Tech of 2026 Awards (published Fri Nov 6, 10:30 AM ET) picks Hark in the AI section, writing that it "rose to the top for a simple reason: it’s beautiful" and calling the interface "pretty, intuitive, and anticipating." Hark's company account posted at 2:08 PM ET: "Thank you @Gizmodo for naming Hark Pro the Best AI Software of 2026! (We agree)" (company post). Brett Adcock, Hark's founder, quoted that post with "LFG" (founder post). No product change shipped with the award. Reopened Monday; unchanged.

• A user says Hark filed a flight-delay claim worth €1,200. Gokul Madan (@Gokul_Madan) wrote at 4:31 AM ET Monday that Hark "found me €1,200 for a flight that was delayed 5 hours" and "handled the entire process from filing out the forms and submitting it" (user report). He says it spotted the delay about 3.5 hours before the airline posted. We have not seen the claim or payout. Our earlier Handoff story covered a similar train-refund use.

• Dale Thomas says Hark caught a 26-hour layover. Dale Thomas (@dalethomas50), whose hotel and flight catches we covered Friday, wrote at 2:55 PM ET Saturday that Hark was helping plan a trip to Japan for 5 people. He says a connection he booked arrived at 5 PM, and "Hark reviewed the trip and immediately noticed the connecting flight was at 7 PM the following day, meaning we would have a 26-hour layover at the connection." (user report) We have not seen his bookings.

• A user says Hark cleaned up an old X profile. Suneet Singh (@suneet21) wrote at 12:47 AM ET Saturday: "Hadn’t touched my X profile in years. Asked @hark_labs to clean it up and it just… did it. New bio, new photo, and a few people it thought I should talk to." (user report) Confirm before you let Hark change public profiles. Reopened Monday; unchanged.

• Adcock says CUA agent traffic "looks human." Replying overnight to a thread on whether platforms undercount agent orders, Brett Adcock, Hark's founder, wrote at 11:17 PM ET Friday: "There’s no agent attribution via CUA. It looks human" (founder post). Separately at 3:53 PM ET he told a user asking for wife/family pooled usage: "We don’t have this feature right now but will discuss it" (founder post; not a ship). Reopened Monday; no family feature ship; replies unchanged.

• A user asks whether Hark will get voice mode. Earlier Saturday, at 8:39 AM ET, Thomas asked Hark: "any chance voice mode, like Grok Bot has, is on the roadmap?" (user report) We found no reply from Hark or Brett Adcock, Hark's founder, in that thread as of Sunday morning.

Known issues:

• A user says mobile attachments won't download; Adcock replied. Gsspai (@Gsspai) wrote at 11:02 PM ET Sunday: "On mobile, though, I still can't download attachments. I already emailed support about this, but I haven't heard back" (user report). Adcock replied at 11:42 PM ET: "You should be able to long press and download. Will take a look" (founder post). We did not reproduce this, and we have not seen a fix.

• Two more users report limit and CAPTCHA trouble. Lê Tuấn Linh (@LeTuanLinh_me) wrote at 4:24 AM ET Monday that right after a monthly reset he was "instantly locked out with a 'reached daily limit' error despite 0 usage" (user report, screenshots not checked). A user posting as "-" (@goodassstuff) asked Adcock at 1:10 AM ET to improve CAPTCHA solving, saying it "freezes up, fails to select the right tiles, or does it so slowly the CAPTCHA times out" (user report). We reproduced neither. See our Dec 1 reset story.

• A free-plan user says one chat used the month. Ang Dae Naray (@AAnggapra50811) wrote at 11:35 PM ET Saturday that after a Pro trial ended, "after just one chat it said I'd hit my usage limit and have to wait until next month." The post asks whether that is a bug (user report). We did not reproduce it. Hark's Using Hark page says the free plan may "reach a usage cap" without giving amounts. See our Dec 1 reset story.

• Docs describe auto top-up; a user says buy-credits failed. Hark's Using Hark page still says that with auto top-up, "Hark will replenish your credits with an amount you decide, up to a ceiling you predetermine." fred bloggs (@llivloc123) wrote at 5:36 AM ET Saturday that he could not buy extra credits despite the site saying he could (user report). No reply in the checked thread through Monday morning. We did not reproduce. Pricing still lists Free / Pro² $20 / Pro³ $100 for more usage.

Last checked: Nov 9, 2026, 5:56 to 6:58 AM ET across the four desks; weekend items were reopened Monday morning and are unchanged unless the line says so.

What changed: Monday issue covering Saturday through Monday morning. Weekend digest lines from Nov 7 and Nov 8 were refreshed Monday rather than published as separate weekend issues.

Coming up

From Friday's week in review:

Shipped items awaiting our coverage, plus announcements, hints and scheduled changes. Anything marked unverified has no company date.

• Cursor Voice Chat with Agents (shipped Nov 6; our story Monday). Live in the Nov 6 changelog. Our story is dated Monday Nov 9 and is not on the site yet. Watch: Monday's issue, and the docs.
• Grok Bot 0.79.x builds on the download server (unverified as a release). HEAD requests for 0.79.0, 0.79.1 and 0.79.2 darwin-arm64 returned 200 with last-modified Nov 5 UTC; the changelog still lists 0.78.1 as latest. Watch: a 0.79 changelog entry.
• Grok Bot 1.0.0 (rumor). Kevin Logan's earlier claim still has no company date. A 1.0.0 macOS installer path still answered 200 on Friday, dated Oct 27. Watch: a 1.0.0 changelog entry.
• Phone calls in the changelog and docs (expected). Official on X, missing from x.ai/changelog/bot. Watch: plan, country and usage notes.
• Any Bot in Slack (staff hint, unverified). Lauren Tan, whose bio reads "grok @bot at @SpaceXSI", wrote that "soon you'll be able to add any bot to slack". No follow-up ship post in this week's @bot or @poteto originals we read. Watch: the changelog.
• A Cursor 3.28 fix for Windows (staff, no date). Staff say a fix is identified; for now they advise 3.27. Watch: a 3.28 patch or a staff "fixed" reply on the thread.
• Two tracked Cursor bugs (staff, no timeline). deanrie on subagent effort: subagents follow the model picker's saved effort, with no timeline. Markdown Preview links with line anchors still fail; deanrie wrote the issue is tracked. Watch: staff replies on both threads.
• ChatGPT and Codex shared usage (scheduled). Takes effect in December (story). Watch: Help Center updates as the date nears.
• Grok Bot for sales ops playbook (planned Monday, Nov 9). SpaceXSI's sales team described splitting its deal cycle across shared Bots (Account Intel, Deal, Quote). We held the playbook for Monday's issue. Watch: Monday's issue, not a release date for a new feature.
• Anthropic Residency applications close Monday, Nov 23, at 2:59 AM ET. That is Sunday, Nov 22, at 11:59 PM PT, as listed in the morning issue. Paid research placements of six weeks to six months (announcement, US posting). Watch: the application deadline.
• GPT-6 Cyber expanding beyond the first Daybreak Red cohort (no date, unverified). OpenAI says availability expands over time (announcement). Watch: an availability update from OpenAI.
• Pierre acquisition close (no date, unverified). OpenAI says closing needs regulatory approvals (announcement). Watch: a closing announcement, then post-close Codex product notes.
• Later (scheduled or confirmed). Nov 10: "How a Series A Startup Runs on Grok Bot" at 1:00 PM ET (from the workshops page's 6:00 PM UTC). Also listed: Nov 12 Shared Bots, Nov 13 Coinbase on Grok Bot, Nov 16 Grok Bot 101. Nov 16: Claude Enterprise turns on merged chat and Cowork by default (story). Nov 20: OpenAI's updated Usage Policies take effect (story).

Last week's watch list, checked. Built-in calls: shipped as an official @bot announcement; still missing from the changelog (see above). Grok Bot 1.0.0: still rumor; file still on the download server. Any Bot in Slack: hint unchanged, no ship. Cursor 3.28 Windows fix: still open, downgrade advice stands. Subagent effort and Markdown Preview anchors: still open, no timeline. Claude pull-down search: no App Store ship we saw. ChatGPT Chat and Work together: Tibo's earlier hint, no company date this week. Hark Mac release notes: still not published by Hark. Anthropic S-1 price range: no amendment we found this week. Codex x TwitchCon: entries closed Nov 2; our story was updated past tense.

Last checked: Nov 6, 2026, 4:20 to 4:48 PM ET (Grok Bot and Cursor changelogs, status summary and incidents, staff X posts, Cursor forum threads, Cursor Workshops, marketplace Featured row, download server HEADs, Anthropic Residency pages, the SpaceXSI sales post, OpenAI's GPT-6 Cyber and Pierre posts).

Daily Bot Check

1. How is Cursor's new voice chat different from dictation?
   • a) It is a two-way call where the agent answers out loud while coding runs
   • b) It only turns your speech into text in the chat input
   • c) It works only in Cursor for iOS
2. What does a ChatGPT dot ask before filling an emailed sign-in code?
   • a) Nothing, it always fills codes silently
   • b) Whether to allow it once, always, or cancel
   • c) For your email password
3. What do Cursor staff suggest for Windows users whose agent won't start on 3.28 or 3.29?
   • a) Turn off HTTP/2
   • b) Reset the Bot's computer
   • c) Install 3.27 and set update mode to manual

Answers: 1) a, 2) b, 3) c

You're getting this because you subscribed to The Daily Bot. Reply to this email with corrections or tips; a person reads every reply. The website is temporarily offline; fuller articles will keep arriving in this email until it is back.

How we label: confirmed means we checked a primary source or a staff post; reported means a credible claim we haven't verified; "Our read" is our own analysis.