If you make short promo or channel videos, six Hark users say it scripted, voiced, scored or edited theirs from a chat, though Hark hasn't announced video. Also inside: OpenAI says it refunded credits drawn in the OpenCode Console attack, users are giving Grok Bot a phone line through a calling service, and staff now have a workaround for local tools stuck on "Never." Since it's Monday, Friday's week in review comes first, and further down are 19 stories that went up after Friday's email.

In this issue: The week in review · Hark video · What's new (4) · Since our last email (19) · How people use it (3) · Workflows (2) · Digest (21) · Known issues (7) · Coming up · Daily Bot Check

The week in review

From Friday's week in review, Oct 12 to 16, which went up after Friday's email.

• New in Grok Bot: a faster app with a new look and six named connectors (Oct 16), X Chat reading (Oct 16), Bot email aliases (story), plugin submissions (story) and Grok Voice 3.0 (story).
• New plan: new customers now get Grok Bot from the $30 Premium tier of The X Subscription (SpaceXSI). Existing Cursor and SuperGrok plans keep working.
• Broke, then recovered: routines were held back Tuesday morning and event triggers failed that afternoon (status); Bot computers were unreachable from about 9 AM to 5 PM ET Wednesday (status).
• What to do now: update the app, and check which of last week's routines actually ran, since Tuesday's held runs weren't replayed.

Opinion: Grok Bot should turn Musk's purchase pledge into a receipt ledger argues for a Grok Bot purchase ledger with a claim button.

Today's lead

Hark users say it makes whole videos from a chat (reported, user reports)

What happened: Six people have posted videos they say Hark made or helped make, from animation and voiceover to editing and uploading. Uzer Kayat posted Sunday that both of his app launch videos "were made with @hark_labs," and that "Hark handled the animation, the voiceover and the background music." A developer who goes by olax_bonnke posted that he made "a full video ad" for a gas plant app with "Script, a Nigerian voiceover, captions, music," and that when he wrote "the voice sounds flat," Hark "re-recorded it line by line." enrico said "Hark did the title, descriptions etc and upload" for a video.

Who's affected: Hark Pro users who make short videos. These are user reports. Hark hasn't announced video, and its pricing page lists "Orders food," "Juggles email" and "Creates docs," with nothing about video.

Why it matters: Our read: a short promo usually takes a script tool, a voice tool, an editor and a music library. These users describe doing it in one chat, which helps if you ship small apps or run a channel alone. Quality varies, though: Charlie Shenmi wrote that most of a one-minute TV drama edit "felt rather chaotic," with "shots repeating constantly." Iterating costs usage too. Kayat said he went through "at least 6 to 8 versions," and Brett Adcock, who founded Hark, replied to quota complaints: "Good feedback, will work on it".

What to do (these steps come from what the users describe; we haven't tried them):

1. Give Hark the raw material: app screenshots, your own footage, or permission to find clips.
2. Say what you want: length, style, voiceover, music.
3. Give specific notes on each draft, like "the voice sounds flat."
4. Ask for a second cut if you want one, like a meme version.
5. Check you have the rights to every clip and song before you post.

Prompt (ours, untested):

Make a 45-second animated launch video for my app from these screenshots. Write the script, add a voiceover and background music, and show me a draft before you change anything big. Only use footage and music I'm allowed to post.

What to watch: whether Hark adds video to its Using Hark page or names the model, and whether Adcock changes quotas after this week's feedback.

Read the full story

What's new

No new Grok Bot version: the changelog still ends at 0.70.2 from Oct 11, and Cursor's status page is All Systems Operational this morning. (confirmed)

OpenAI refunds credits drawn in the OpenCode Console attack (OpenAI · confirmed, staff)

If you saved an OpenAI API key in OpenCode Console, OpenAI says credits drawn in the attack were refunded. Tibo Sottiaux, who works on Codex and ChatGPT, posted at 12:12 AM ET today that OpenAI worked with OpenCode to get affected API customers "back in business and any credits refunded." OpenCode said the breach "impacted 1.4% of OpenCode Console users". Still delete the key and set a hard spend limit. Read the full story

Your Bot can make phone calls, with a calling service (Grok Bot · reported)

Brian Roemmele posted Sunday that his Bot booked a restaurant table by phone and "waited for my OK before it dialed," and shared the template. Grok Bot has no number of its own; ClawCall's site says it "currently handles English-language calls to US +1 numbers," with paid plans from $4.99 a month. To try it, message a Bot "Get the ClawCall MCP so you can make phone calls for me," authorize the card, and ask for one call with a read-back first. We didn't place a call. Read the full story

Claude Personal Finance spotted, not confirmed (Anthropic · rumor)

TestingCatalog posted screens Sunday of an unreleased Claude "Personal Finance" feature that links banks through Plaid: "Read-only, so Claude can't move money." Anthropic hasn't confirmed it, and the leaked screen says "Plaid isn't available on this account." Read the full story

Local tools stuck on "Never": staff now have a workaround (Grok Bot · confirmed, staff; not fixed)

Colin explained Sunday that "for some newer accounts, the Grok Bot app doesn't finish linking your computer to your account," so it's treated as Never "no matter what Settings says." On grok.com, open Settings, then Privacy, click Edit and save, wait about a minute and ask again; if it still refuses, quit and reopen Grok Bot. It's tracked, and one user still reported it at 12:31 AM ET Monday. Read today's issue

Since our last email

These went up on the site after Friday's 8:22 AM ET email.

Grok Bot

• Grok Bot update: 3x faster computer tasks, new look, new connectors (release, confirmed). The official @bot account says "Computer tasks finish 3x sooner on average" and lists Robinhood, Coinbase, eToro, Trello, PostHog and beehiiv connectors. Update the app, and set an Auto Review rule that asks you before any order if you add a trading connector. Story
• Grok Bots can now read your X Chat messages and requests (feature, confirmed). Paul Vann, a SpaceXSI staff member, posted that Bot access to "your @chat messages and message requests" is "now live." Connect the X plugin and give your Bot clear rules first. Story
• Gmail connector still turns your email links into Google redirects (known issue, confirmed). Colin said the fix "isn't available to your accounts yet." Until then, re-paste links yourself in Gmail before you send. Story
• Grok Bot for home services: a daily call review and booking rate (playbook, reported). An Oklahoma HVAC and plumbing owner described a Bot that sorts every inbound call at 5 PM and reports a booking rate by trade. We haven't run it; the playbook has a routine you can copy. Story
• Give your Bots house rules so they don't freeze their computer (tip, confirmed, staff). deanrie posted a rules file that makes Bots check memory and run one heavy job at a time. Save it as /workspace/RESOURCE_RULES.md, tell each Bot to follow it, and if a computer freezes, don't click Update, Recover or Reset. Story

Cursor

• Cloud Agents your Bot starts can inherit Grok 4.7 Fast (tip, confirmed, staff). deanrie found one account's Cloud Agents default "is saved as Grok 4.7 with Fast turned on," so Bot-launched agents ran as Grok 4.7 High Fast. Check the Default Model at cursor.com/dashboard and turn off Fast. Story
• Stop Cursor's agent from running drive-wide deletes (tip, confirmed issue, one report). One user reports an agent in Run Everything mode deleted "Almost the entire D: drive" on Windows, and staff say it's tracked. Switch to Allowlist, turn on File-Deletion and External-File Protection, and add a beforeShellExecution hook that blocks recursive deletes. Story
• Cursor's /review-bugbot subagent always bills as Composer 2.5 Fast (known issue, confirmed). kevinn confirmed it "is currently set to run on Composer 2.5 Fast" whatever model you pick. Review in the main chat with no subagents, or use your own subagent with model: composer-2.5[fast=false]. Story
• Cursor CLI 2026.10.16 fixes the Windows worker crash (fix, confirmed). deanrie said the "Waiting for a worker" crash "is fixed in Cursor Agent CLI version 2026.10.16 and newer." Update the CLI and install Git for Windows. Story
• Cursor cloud agents can now build shareable canvases (feature, confirmed). The changelog says cloud agents "can now build interactive canvases," which you can share as a link. A beta Environments API lets you script agent environments. Story

Anthropic

• Anthropic says agents its staff direct now log six times their hours (explainer, confirmed). Anthropic's report says by August "the agents worked 6 times as many hours as their human handlers." Our read: people reading agent output are now the bottleneck. Story
• Anthropic plans to put 5% of its shares in a new public benefit fund (explainer, confirmed). Subject to stockholder approval, Anthropic says 5% of its shares "will be committed today," with more possible as its value grows. The post describes no change for Claude users or prices. Story
• Anthropic's draft AI bill, as reported, next to its June framework (explainer, reported). Politico reported that Anthropic pitched lawmakers a draft bill under which only the Commerce Department, through a federal judge, could halt a model. Anthropic hasn't posted the bill (Reuters). Story
• Claude finds NIST ECDSA curve seeds came from Unix timestamps (explainer, confirmed). Anthropic says a research model traced the NIST curve seeds to the lrand48 generator and 1990s timestamps. Its own line: "There is no action to be taken or things to do differently." Story

OpenAI

• GPT-6 Pro in ChatGPT now splits tasks across parallel agents (release, confirmed). @OpenAI posted Friday that "multiple agents can tackle different parts of a task in parallel" for Pro users in Chat. Open each finished piece to check it before trusting the combined answer. Story
• ChatGPT Slides: build a team deck in Space, then export it (feature, confirmed). OpenAI's help article says Slides is on Pro, Business and Enterprise plans with Space access. Ask Work for a deck, edit it with teammates, then present or export. Story
• GPT-Live voices: describe one in text, no recording needed (release, confirmed). OpenAI's API changelog says prompt-to-voice is out for all API customers, and creating a voice is free. Send a description to /v1/audio/voices and pass the voice ID to a gpt-live-1 session. Story
• ChatGPT Chrome tabs: Chat reads them, Work controls them (release, confirmed). @ChatGPT said Chat can look across your open tabs and Work can fill forms, send messages, download files and group tabs. Install the ChatGPT Chrome extension to use it. Story

Hark

• Hark says its free pizza night ran 203,500 Handoff browser sessions (explainer, reported, founder's claims). Adcock posted that "Over 203,500 browser sessions were run via our model Handoff" during Friday's giveaway. If you connected Gmail or Outlook for the pizza, you can request an export of what Hark holds under System Preferences, then Data Management. Story

How people are using it

• A personal trainer from your watch and mattress data (Grok Bot). Paul Vann, whose X bio lists @spacexsi and the X API, posted at 3:36 PM ET Sunday that he "asked my @bot to connect my @Garmin and @eightsleep data, and operate as my personal trainer." He says it can "triage my sleep data between my watch and @eightsleep." He didn't say which connectors he used. (reported, one user who works at the company)
• One home for everything a Bot writes (Grok Bot). Melvin Vivas, who shares Grok Bot templates, described at 11:21 PM ET Sunday how he files Bot output: notes to Notion as markdown, documents to Google Docs, Sheets and Slides, and "anything else" to Google Drive, so it's "ready for sharing." (reported, one user)
• A Bot made from inside the Grok app (Grok Bot). Carolina Dill wrote at 5:34 PM ET Sunday that while she chatted with Grok, it "offered to create a Grok Bot for it," and a few seconds later the Bot was in her list "already set up with its own folder and a short README." We haven't seen this documented. (reported, one user)

Workflows and Setup of the day

Setup of the day: a crypto creator's bot prints a fact-checked one-page news sheet each morning (Grok Bot · self-reported)

James Rule, a crypto creator, says "I run a Grok Bot I call my Chief of Research," and estimates it saves "about an hour and a half a day." In his words:

1. "Every morning it pulls the last 24 hours of crypto news."
2. It "checks it against primary sources."
3. It "builds my one-page briefing."
4. It "prints it on my printer every morning at 6:00am Central" (second post).
5. "It also preps and handles my posts once I approve them."

He also promotes his free bot templates, the result is self-reported, and we have not run it. Read the full story

A checker bot that looks for the output, not the green run (Grok Bot · self-reported, one user)

@grovr_au keeps one bot that "checks whether a routine actually did its job, not just whether it says it ran," and says it caught a sync that had "succeeded" for six days without writing anything. It "checks where the output is expected to see if it landed," they replied. One user's report, and we have not run it. Read the full story

The digest

Grok Bot

• Routines can lose your Mac after an update. Follow-up to our Oct 17 item. One user's Mac "registered as a new computer" on the first launch after the update, so a routine kept pointing at the old one, mohitjain said at 5:12 AM ET. It's tracked. Staff tip: name the computer in the routine, like "If the saved computer isn't connected, use the connected computer with that same name." (confirmed, staff; one report)
• Bots missing after the X Subscription switch? Don't reset. For one user, "The switch to The X Subscription didn't fully finish combining your accounts," mohitjain wrote at 4:41 AM ET. The Bots, chats and files "aren't lost," and "Retrying, Reset, or creating a new Bot won't bring them back." Wait for staff. (confirmed, staff; one report)
• Windows: let updates finish before you shut down. A downloaded update installs when the app quits, and "if it's interrupted you're left with a partial or empty folder," Colin said at 11:46 AM ET Sunday. Turning off Automatic Updates doesn't stop that. Click Restart to finish installing while you're at the PC, and wait until Settings shows the new version before shutting down or sleeping. If the folder is already empty, reinstall; "your Bots and settings are kept." (confirmed, staff; one report)
• iPhone app can't show connector cards yet. On iOS, cards from connectors such as Fastmail show an update message even when you're up to date. Colin said at 11:58 AM ET Sunday that support "is already built and will come in an upcoming iOS update." Meanwhile, open the chat on your Mac or ask the Bot to write the details as a regular message. (confirmed, staff)
• X Subscription tiers, in staff's words. mohitjain wrote at 3:59 AM ET that "Pro ($20) maps to Premium ($30), Pro+ ($60) to Super ($100)," and usage "is now one weekly pool shared across Cursor and the Grok apps." Tip from the same post: "Fast models use the pool about twice as fast." In Korea, plans have fixed won prices, and ₩45,000 includes VAT. (confirmed, staff) Background story
• Teams with one paid seat can't change tiers yet. On The X Subscription, changing a seat's tier "isn't possible from the dashboard right now when the team has a single paid seat," mohitjain said at 4:03 AM ET. Email [email protected] with the team name, the exact change and a link to the thread. (confirmed, staff)
• Annual billing is for individuals only. Ultra is "$2,000 per person per year," and "Team plans (one bill + admin controls) are monthly only today," mohitjain said at 5:01 AM ET. (confirmed, staff)
• Template installs failing? Check your X account. Monetization Coach posted at 12:01 AM ET Monday that when bot template installs failed, "Usually, it's because Grok Bot is connected to a different X account than the one I'm logged into for X." (reported, one user)
• A formal verification skill in pstack. Lauren (@poteto), whose X bio reads "Grok @Bot at @SpaceXSI", posted at 5:05 PM ET Sunday that pstack v0.15.27 "ships with a new skill /create-formal-verification-skill," a beta "meta-skill that creates a new skill and formal verification suite." (confirmed, author)

Cursor

• Measure how much of an agent's patch survives review. Cursor staff member deanrie suggested Plan mode for fixes and refactors and a "reuse first" rule in Project Rules, noting "rules are a strong hint, not a guarantee." To compare models, he added at 6:04 AM ET: in the cloud SDK "Each run through Agent.create gets its own branch," found in result.git.branches, so git diff --stat <agent branch> <final branch> shows the cleanup. Locally, commit the agent's output first as its own commit. (confirmed, staff)
• Anthropic's cyber filter blocks a chat? Start fresh or switch models. The block "comes from Anthropic's own cyber-content filter, not Cursor," and it reads the whole conversation, mohitjain said at 4:41 AM ET. Start a new agent, or use a non-Anthropic model for security work. Anthropic's Cyber Verification Program relaxes the filter, and "it's currently available on Enterprise plans." (confirmed, staff)
• Opus 5.5 over your own AWS Bedrock key can break long chats. It's "an issue we're tracking," mohitjain said at 4:01 AM ET. Start a new chat, or switch a stuck one to "Opus 5 or 4.6." (confirmed, staff)
• The SDK's systemPrompt isn't open to outside accounts. mohitjain said at 4:00 AM ET it "isn't open to external accounts yet," and the "unknown option" error is misleading. Use an always-applied project rule or AGENTS.md instead. (confirmed, staff)
• Too many Agents windows on Windows launch. Set Cursor Settings, General, Startup, Window Restoration to Last Used Windows, close the extras and quit fully, mohitjain said at 3:58 AM ET. "It's an issue we're tracking." (confirmed, staff)
• Don't remove an empty duplicate Agents section. Both entries "point at the same folder," so removing one hides your chats, Colin said at 11:47 AM ET Sunday. To get them back, use Open Workspace and pick the same folder. (confirmed, staff)
• CLI resume fails when /tmp is locked down. Persistent sessions use "a fixed /tmp path rather than respecting $TMPDIR," Colin said at 11:49 AM ET Sunday. Workaround: an admin creates /tmp/cursor-agent-persist-<your uid> owned by you with chmod 700. (confirmed, staff)
• Remote server freezing? Don't open the whole disk. An SSH window with / open can push a small server to its memory limit, Colin said at 11:43 AM ET Sunday. Open only your project folder. .cursorignore "won't lower memory use there." (confirmed, staff)
• No-Git plugin uploads aren't coming soon. Colin said at 6:24 AM ET that manual plugin uploads aren't "on our short-term roadmap," but admins can create Team Rules and Team Hooks in the web dashboard, which "sync to all team members automatically." (confirmed, staff)

Anthropic

• Claude Code 2.1.303: a bug-fix release. Version 2.1.303 reached npm at 5:15 PM ET Sunday and is the latest and next tag; stable stays on 2.1.292. The CHANGELOG says only "Bug fixes and reliability improvements". The Agent SDK moved to 0.3.303 the same minute. No need to rush the update. (confirmed)

OpenAI

• Edit earlier messages in a thread. OpenAI's Tibo Sottiaux, who works on Codex and ChatGPT, posted at 9:18 PM ET Sunday under "Day 14/": "Small quality of life improvement. You can now edit previous messages within a thread in the desktop app. Long time requested, and took way too long to get that one out." In a reply, he wrote "Registers as time travel". BlockBeats reports it as a Codex desktop feature. Tibo named no plans or versions, and OpenAI has no docs or changelog entry for it yet. Our suggestion: when a requirement several turns back was wrong, fix it there instead of stacking corrections. (confirmed, staff)

Hark

• Adcock says active users rose 67%. Brett Adcock, Hark's founder, posted at 1:00 AM ET Monday that "active users are up +67% over last weekend." He gave no base number, and nobody outside Hark has checked it. In a follow-up he said "the free tier is really generous". (reported, founder's claim, unverified)
• A company demo: a DMV address change. Lara Lu, whose X bio reads "mobile @hark_labs", posted at 2:24 PM ET Sunday that Hark "signed into the DMV, found my old address, and filed the change. 2 min." She appears to work at Hark, so treat it as a company demo, not an independent review. (company demo)

Known issues and workarounds

The newest open items from Known issues, plus two touched today:

• /review-bugbot bills as Composer 2.5 Fast (Cursor, confirmed): review in the main chat with no subagents, or use your own subagent set to [fast=false]. Details
• Gmail connector rewrites links into Google redirects (Grok Bot, confirmed, fix in progress): add or re-paste links yourself in Gmail before sending. Details
• Old Grok app messages keep reaching some Bots (Grok Bot, confirmed): delete old Grok Tasks that point at your Bots; don't just archive them. Details
• Skill sync can overwrite your local skills folder (Cursor, confirmed): keep Sync Skills for Cloud Agents off for now, and check the Trash if it hit you. Details
• Amazon shows Grok Bots an unauthorized-agent page (Grok Bot, reported): Musk says shop elsewhere, and Amazon's Conditions of Use forbid working around it. Details
• Local tools stuck on "Never" (Grok Bot, confirmed, tracked; touched today): on grok.com, open Settings, then Privacy, click Edit and save, wait a minute, then quit and reopen Grok Bot if it still refuses. Forum
• A Mac re-registered after an update breaks routines tied to it (Grok Bot, confirmed, tracked; touched today): name the computer in the routine so it can fall back to the connected computer with the same name. Forum

Coming up

From Friday's week in review and today's issue. Everything here is announced, hinted or scheduled, not shipped.

On the calendar

• Oct 20 and 22: OpenAI DevDay Exchange. OpenAI lists events in Tokyo on Oct 20 and Seoul on Oct 22, by application (event page). (scheduled)
• Oct 23: OpenAI API shutdowns. gpt-3.5-turbo, gpt-4, o1, o3-mini, o4-mini and others shut down that day (deprecations). If a Bot or script calls one, switch it this week. (scheduled)

Confirmed, no date

• A prompt link that opens a Bot with your text typed in. Colin said it's coming in an upcoming update, and "The message is filled in but never sent automatically." (story)
• Connector cards on iPhone. Colin said support "is already built and will come in an upcoming iOS update."
• Fixes in flight. The Gmail link fix is being built; the skill sync overwrite was flagged to the team; the Grok app replay, skipped X reply tags, the shared webhook limit, local tools stuck on Never and the re-registered Mac are tracked issues.

Early signals (unverified)

• Saved web passwords for Bots. Musk replied "Coming very soon." A hint, not a date.
• @bot tagging outside the US. Lauren Tan wrote "we're not in EU/UK/Australia yet, coming soon." (staff hint)
• More after Friday's update. Parker Conrad wrote "More soon." (staff hint)
• Written terms for Musk's purchase pledge. Watch for a claim form or help page behind the pledge. (open question)
• Claude Personal Finance. TestingCatalog's sighting of Plaid bank linking is unconfirmed by Anthropic. (rumor)

Daily Bot Check

1. Has Hark announced a video feature? (a) Yes, on its pricing page (b) No, these are user reports (c) Yes, in the Pro 2.4.0 notes
2. What does Grok Bot need before it can phone a restaurant? (a) Nothing, it has its own number (b) A calling service such as ClawCall or Bland (c) A Cursor Ultra plan
3. Staff's workaround when local tools wrongly say Never starts where? (a) Reinstall Grok Bot (b) Settings, then Privacy on grok.com, then Save (c) Reset the Bot's computer

Answers: 1. (b) No, these are user reports. 2. (b) A calling service such as ClawCall or Bland. 3. (b) Settings, then Privacy on grok.com, then Save.

Every story in this email

• Hark users say it makes whole videos from a chat
• OpenAI refunds API credits drawn in the OpenCode Console attack
• Send your Bot to make the phone calls you keep putting off
• Unreleased Claude screens show Personal Finance with Plaid bank linking
• A crypto creator's bot prints a fact-checked one-page news sheet each morning
• A checker bot that looks for the output, not the green run
• Grok Bot this week: a faster app, X Chat, a new plan and two outages
• Grok Bot should turn Musk's purchase pledge into a receipt ledger
• Grok Bot update: 3x faster computer tasks, new look, new connectors
• Grok Bots can now read your X Chat messages and requests
• Gmail connector still turns your email links into Google redirects
• Grok Bot for home services: a daily call review and booking rate
• Give your Bots house rules so they don't freeze their computer
• Cloud Agents your Bot starts can inherit Grok 4.7 Fast
• Stop Cursor's agent from running drive-wide deletes
• Cursor's /review-bugbot subagent always bills as Composer 2.5 Fast
• Cursor CLI 2026.10.16 fixes the Windows worker crash
• Cursor cloud agents can now build shareable canvases
• Anthropic says agents its staff direct now log six times their hours
• Anthropic plans to put 5% of its shares in a new public benefit fund
• Anthropic's draft AI bill, as reported, next to its June framework
• Claude finds NIST ECDSA curve seeds came from Unix timestamps
• GPT-6 Pro in ChatGPT now splits tasks across parallel agents
• ChatGPT Slides: build a team deck in Space, then export it
• GPT-Live voices: describe one in text, no recording needed
• ChatGPT Chrome tabs: Chat reads them, Work controls them
• Hark says its free pizza night ran 203,500 Handoff browser sessions
• Grok Bot now starts at $30 on The X Subscription
• Full issue, Oct 19

You're getting this because you subscribed at thedailybot.vercel.app. Reply with corrections or tips; a person reads every reply.

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.