Pages that update themselves: what people built on night one

Grok BotWhat's new#

Ask your Bot for a dashboard or brief as a page, then let a routine rebuild it. Jenny Wen, whose X bio reads "head of design @cursor_ai @bot," wrote at 5:07 PM ET Wednesday that she has "a few routines set up" and "my bot makes pages for these and updates them on a weekly basis." That answers the open question in our Oct 21 launch story, which we've updated. Overnight, two users posted day-old setups: Carolina Dill's numbers dashboard that "updates every morning" (post) and Cheesecake's chief-of-staff dashboard, with a prompt you can paste (post). Official docs still don't describe how pages refresh. (Grok Bot; reported; staff tip and two single-user reports)

If you're on Claude Pro or Max in the US, Claude Duet's waitlist is still the way in. Anthropic calls it "a new app where Claude works as a proactive agent," and the Help Center says "Invitations go out in waves." Our full guide, from yesterday, covers how to join and what's still unclear. (Anthropic; confirmed)

If you use OpenAI models through a router or AI gateway, ask that provider whether it has OpenAI's approval. From Nov 20, OpenAI's updated Usage Policies say a service "that enables third parties to directly access OpenAI models" must "obtain our prior review and approval," and OpenAI's post names "model routers or AI gateways." The policy adds: "This requirement does not apply solely because an application integrates OpenAI models to power its features or functionality." The Oct 21 update also tightens the wording on using OpenAI outputs to train rival models. OpenAI hasn't published an approval process yet. (OpenAI; confirmed; not legal advice)

If you have Claude API code to move to API version 2026-10-21, update Claude Code and run /claude-api api-version. Claude Code 2.1.307 added the command "to move Claude API code to API version 2026-10-21," per the CHANGELOG, and it reached npm at 4:47 PM ET Wednesday. The same release fixed "HTTP MCP tool calls being sent twice when the connection closed mid-call" and added a claude_code.safety_event telemetry event that the monitoring docs don't describe yet. @ClaudeDevs wrote that "2023-06-01 is not deprecated," so you can move when you're ready. (Anthropic; confirmed)

Before you act on Hark's research, check the exact wording yourself. In field notes on Medium, Lakshmi narayana .U says Hark got a film festival shortlist right ("All three planning calls held up"), then reported a subtitle rule contradiction that wasn't there. A Clermont-Ferrand rule read "non embedded subtitles only," with "non" at the end of a line, and the agent "omitted the word that sat at the line break." The author calls it "one word at a line break" and writes that it is "one run, three festivals." We pulled out the author's four checks and a prompt you can paste. (Hark; reported, one user's report)

Grok BotHow people are using it#

• Thursday showcase: five recurring AI jobs people run, from a pizzeria's 6 AM books report in Claude Code to Grok Bot CRM reports, a team's CI fixes in GitHub Actions, a Hark usage panel and a ChatGPT dot as an assistant, each with a starter prompt. Credit: Chris Simone, Brad Smith, Akos, LarMedAI, Treb Gatte (self-reported; disclosures in the piece).
• Setup of the day: A routine that clears spam texts and junk mail on a timer (Grok Bot). @FoundAt5280 says an hourly routine now handles what used to be "no less then 30 text messages every day that were spam," and @nerdalator's Bot "runs every morning" and asks "who to unsubscribe from." Neither says how the Bot reaches their messages, and we found no official Grok Bot feature for reading texts. (self-reported, two users)
• A 45-minute morning walk call with a ChatGPT dot (OpenAI). Riley Brown, a creator whose bio reads "Building an @agentnative_ company," wrote at 9:35 AM ET Wednesday: "Every morning I call my dot and put my AirPods in and I walk have a conversation for about 45 minutes." Riley answers email, writes instructions for editors and hands longer jobs to Codex, and described the same call on Oct 1. Dots need ChatGPT Pro or Business Premium, per OpenAI's Help Center. (self-reported, one user)
• A solo founder's weekday drafts, with selling kept human (Grok Bot). @bageventio, whose bio reads "Founder of BagEvent (since 2015) & MarketUP," wrote at 4:48 AM ET Tuesday: "Every weekday it drafts my X and LinkedIn posts, digs up event organizers worth talking to, and makes a 60 second industry news video every Monday." The line they draw: "Talking to customers is on me." In a reply, @KraJan_303 said merging plans "wiped out nine bots on one account." That's unconfirmed; in the similar X Subscription switch cases, staff say the Bots are safe but can land on a different account (our known issue). (self-reported, one user)

Grok BotThe digest#

Grok Bot#

• Your Bot's computer sleeps when nothing is using it, and cron jobs inside it freeze too. Follow-up to our Oct 11 story. Cursor staff member Colin explained at 7:42 AM ET Oct 21 that programs inside the computer, such as a sync, a watchdog, an rclone mount and cron, "don’t count as activity, so they freeze along with it," and "There’s no way to keep the computer running 24/7 today." His advice: "Move overnight work into scheduled Bots rather than cron jobs or daemons inside the computer," make long jobs catch up on wake, and note that "A command a Bot starts in its own shell should keep the computer awake for a while (up to about 30 minutes)." He says there's "no need to Reset, Recover or Update." The poster replied that it "helps me rethink how I’m using my Grok Bot setup," and a second user says their Bot's computer also sleeps overnight and asked for a way to keep it awake. (Grok Bot; confirmed by staff)
• A CI routine on one pull request won't fire if Cursor cloud agents pushed the commits. Colin listed four setups where a PR-scoped CI routine never fires: a PR "opened from a fork," CI that "reports results as commit statuses," a commit in "more than one open pull request," and commits "pushed by the bot through the Cursor GitHub App rather than by you." That last one "is an issue the team is already tracking," he wrote at 3:48 AM ET. Staff workarounds in the same reply: a routine on "PR pushed" that waits for the checks, or "A short scheduled routine that checks the PR’s checks and deletes itself once they finish." After the poster confirmed it was their case, Colin added at 4:44 AM ET that he'd added the report "so the Grok Bot routine side is covered too." (Grok Bot routines and Cursor cloud agents; known issue)

OpenAI#

• Codex for OSS grants move up to Pro 200. Charlie Marsh, whose X bio reads "@OpenAI. Building Ruff, uv, ty, and other high-performance Python tools with the @astral_sh team," posted at 1:37 PM ET Oct 21: "we’re upgrading all Codex for OSS grants to Pro 200!" He said there will "still be the 10,000 grants in this round, now at the higher tier," that past recipients should see the upgrade "automatically," and that "Applications are still open." @OpenAIDevs reposted it. The September round offered "$100 Pro plans," per his Sep 15 post. Disclosure: Marsh works at OpenAI. (confirmed, staff)
• API admins can limit which service tiers a project uses. Per OpenAI's API changelog for Oct 21: "You can now choose which service tiers projects can use in Platform settings. Organization settings can restrict service tiers for all projects." Our read: it's a way to keep a test project off a tier such as Ultrafast. The entry links OpenAI's service tier errors page for what a blocked request returns. (confirmed)
• Tibo: "it’s reset day today." OpenAI's Tibo Sottiaux, whose X bio reads "Codex & ChatGPT @OpenAI," posted at 2:55 AM ET Oct 22, as Day 17 of his daily posts: "Bet you thought we forgot, but no, it’s reset day today." That's his claim; he didn't say which usage limits reset or for whom. In the same post he urged readers to "get Daybreak Blue access, it only takes 15 mins." Daybreak Blue is one of two access tiers "for approved defenders," per the API changelog. (staff claim)

Known issues on OpenAI's status page, Oct 21 to 22 (times ET):

• Oct 21, 10:18 PM to Oct 22, 4:50 AM: Elevated ChatGPT conversation and Responses API errors. "This is impacting users in India, Japan, and Korea," per the status page. Mitigation applied 10:42 PM. Resolved.
• Oct 21, 9:42 AM to 3:34 PM: Elevated FedRAMP API error rates for Astra and Sol. Resolved at 3:34 PM; yesterday's issue had it still monitoring at 2:50 PM, and we've updated that line.

Cursor#

• Cursor Projects can't use your local MCP servers directly. Colin wrote at 4:43 AM ET that "the Project’s coordinator always runs as a Cloud Agent on Cursor’s servers," so it doesn't load ~/.cursor/mcp.json. Agents it starts on your machine can. His steps: make sure Run on includes your machine, then ask for a local agent by name, for example "Start a local agent on my machine to pull last week’s failed Airflow runs using the airflow MCP tools." (Cursor; staff explanation)
• Bugbot and security subagents failing with a model-not-found error: a fix is rolling out. Colin said at 4:39 AM ET: "A fix is already done and should reach you automatically, without reinstalling." Until it reaches you, ask the main chat to do the review "yourself, without using a subagent." (Cursor; staff says rolling out)
• The Explore subagent model setting only covers Explore subagents. Colin explained at 3:58 AM ET that general-purpose subagents use "the same model as your chat," and the main agent can name a model, a choice that "takes priority over the setting." To steer it, ask in the prompt, for example "launch any subagents with Grok 4.7." (Cursor; staff explanation)
• Cursor windows crash with "oom" at about 4 GB, regardless of RAM. Colin said at 3:45 AM ET it's "a known issue we’re already tracking," worst in long agent chats. His tips: start a new chat regularly, turn off Continue Interrupted Agents under Cursor Settings > General > Startup, and update, since "Newer versions no longer keep full copies of very large files in the chat’s edit history." (Cursor; known issue)

Hark#

• One user reports a sign-in that opened an empty new account. Samuel Voltolini (@Wololand) posted at 2:44 AM ET Thursday that after being logged out, signing back in with a phone number, "the same method used during setup", gave this result: "it shows as a new account, all projects and chats and information gone." Voltolini wrote: "I had the most costly plan when this happened." As of 6:10 AM ET we found no reply from Hark, and status.hark.com listed every component, including Authentication Services, as operational. It is one report so far. (user report)
• iOS rating slips as the app stays on 2.4.0. Apple's App Store listing still shows version 2.4.0 from Oct 16 ("New interactions. Several bug fixes. And more refinements to Hark."). Its average is now 3.74 stars from 337 ratings in Apple's lookup data (the listing rounds it to 3.7), down from 3.85 from 300 when we checked at 2:34 PM ET Wednesday (App Store, 6:09 AM ET Thursday). Google Play still shows its Oct 20 update and 100K+ downloads. (store listings)

Anthropic#

• API platforms can now see when Anthropic restricts one of their end customers. If you send user profiles with your requests, each profile now shows whether Anthropic has an open restriction or ban on that end customer. The release notes for Oct 21 say Anthropic added the enforcement_status and enforcement_actions fields "so you can see whether Anthropic has restricted or banned one of your end customers," plus a new enforcement_status[] filter for listing profiles. The user profiles guide says to send these requests "with no user-profiles- value in the anthropic-beta header, because user-profiles-2026-09-04 and earlier values leave the fields out," and to "Treat an enforcement_status you don't recognize as restricted." On Claude Platform on AWS, "enforcement_status is always unrestricted." At 5:56 AM ET Thursday, the List User Profiles API reference didn't show the new filter yet, so work from the guide. (confirmed)
• Two short Claude incidents, both resolved. Anthropic's status page logged "Degraded performance for Claude Fable 5" from 3:49 PM to 4:29 PM ET Wednesday, opening with "elevated errors on requests to Claude Fable 5," against claude.ai, the Claude API, Claude Code and Claude Cowork. A second incident, "Elevated errors affecting Connectors" on claude.ai and Claude Cowork, opened at 11:50 PM ET Wednesday. Anthropic deployed a fix at 3:28 AM ET and marked it resolved at 4:07 AM ET Thursday. Both were rated minor. Our read: if a connector failed overnight in a scheduled task, run it again. (confirmed)
• Update, Oct 22: Claude Code's stable channel is past the 2.1.304 bug. Our Oct 21 story warned that stable still pointed to 2.1.304, the release with a provider and gateway regression. When we checked npm at 5:56 AM ET Thursday, stable, latest and next all pointed to 2.1.308, which adds only "Bug fixes and reliability improvements," per the CHANGELOG. We've updated the story. (confirmed)