Source: raw/Claude_Cowork_Just_Went_Fully_Hands-Off_Scheduled_Tasks_V2.md (https://www.youtube.com/watch?v=b3OiEUp1GIU, 15:33, uploaded 2026-07-16) · raw/5_Claude_Cowork_Scheduled_Tasks_For_Business_Owners_With_No_Time.md (https://www.youtube.com/watch?v=UfN4l2yCT54, 14:39, uploaded 2026-07-25) · raw/The_Only_Claude_Cowork_Setup_Busy_Entrepreneurs_Actually_Need.md (https://www.youtube.com/watch?v=BMEaJTl3Vo0, 18:22, uploaded 2026-08-16) Creator: Simon Scrapes (YouTube) · Platform: YouTube

Three Simon Scrapes videos from July and August 2026 on the then-new ability to run Claude desktop scheduled tasks in Anthropic’s cloud with the laptop shut. He shows five finance and ops tasks he runs in his own business, builds a subscription audit live, and names the two limits that forced workarounds at the time: no access to local files and no place to store API keys. The method he teaches is writing a clear “done criteria” for each task. All three videos promote his paid Agentic Academy community and offer the task-setup skills as free downloads.

Key Takeaways

  • What changed (July 2026). Before, a business owner had three options: a local scheduled task (runs only while the laptop is awake), Claude Managed Agents (“the developer route”: credential vaults, repos), or a VPS running Claude Code with routines. Now you describe the task, pick a schedule, and it runs on Anthropic’s cloud. He says it draws on normal Pro, Max, Team or Enterprise plan usage rather than separate billing.
  • Create it with Claude. Schedule tab → New task → “Create with Claude” opens a session that interviews you (frequency, approvals, model) and writes the task description. Before the first live run it pulls historical data, runs a trial, and edits its own prompt. The task page lets you change approvals (manual or skip), frequency and model, and each run keeps a log you can open.
  • Loop engineering is the frame. He paraphrases Anthropic’s post: a loop is Claude repeating cycles of work until a done condition is met. Loops start turn-based (you message it), time-based (a schedule) or event-based (an email arrives). An event can be approximated with a schedule, for example checking the inbox every hour.
  • The skill to learn is the done criteria. Each task ends only when a checkable condition holds, for example “every transaction from the week is either matched or put onto the gaps list”, or “check every competitor in the list and say with proof whether they changed anything”.
  • Limit 1: no local files (July 2026). Cloud tasks run in a sandbox and can reach connectors but not folders on your computer. His workaround: move brand voice, templates and other context into Notion or Google Drive and tell the task to read them there. Notion also gives team permissions.
  • Limit 2: no keys. There was no credential vault, and the task would not accept API keys in the prompt, so tasks can use only built-in connectors or OAuth connectors. Two workarounds: a custom remote MCP server (he had Claude adapt a public FreeAgent MCP server, add receipt upload, and host it), or Composio as one MCP connector that holds the credentials for many apps. Encrypted environment variables exist in Managed Agents, which he calls more technical.
  • One Gmail connector, one inbox. The built-in Gmail connector reaches one account. His receipts land in three inboxes, so he connected each Gmail account inside Composio and told the task to search through the Composio MCP.
  • Build time. He says the two accounting automations took about four hours in total, mostly answering questions and reviewing test output, and the subscription audit took about 10 minutes because his FreeAgent connector already worked.
  • It replaced an n8n workflow. His earlier invoice-reconciliation workflow in n8n “stopped working” and he had no time to maintain it. He describes the Claude version as self-maintaining because it reports failures and tries to fix them.

The five tasks

TaskScheduleWhat it doesDone criteria
Invoice and receipt reconciliationFridays 4:00 p.m.Pulls the week’s unexplained transactions from FreeAgent, searches three Gmail inboxes for matching receipts, attaches and explains matches, emails a gaps listEvery transaction is matched or on the gaps list
Subscription creep audit1st of the month, 8:00 a.m.Read-only. Pulls recurring bills, expenses and bank transactions for about eight months; flags price rises, duplicates and anything unchanged for six months; totals this month against lastAn email with the summary, what changed, what looks suspicious, and the totals
Email voice and reply templatesOne-off, then feeds a scheduled inbox taskReads 90 days of sent mail, writes an email voice profile, and turns emails sent five or more times (refunds, contract sends, invoice chases, “the polite no”) into templatesNot stated
Meetings to tasks11:30 a.m. and 4:30 p.m., Monday to FridayPulls Fathom action items and adds them to a Notion task list with owner, due date and a link to the recording momentNot stated
Weekly competitor watchWeeklyChecks about 55 competitors in a Notion table (site, socials, blog), writes top movers and a per-competitor rundown with source linksEvery competitor checked, each marked changed (with proof) or unchanged

Reported results from his own runs:

  • Reconciliation. One run attached 17 receipts from the past week. He had built up about 600 unreconciled transactions over six months before setting it up.
  • Subscription audit. Recurring spend rose about 15% month on month, almost all from Meta ads. It flagged four subscriptions that had crept up (Excalidraw Plus seat, Notion, Slack up 30%, Airtable up 20%), three overlapping AI assistants plus separate API spend, and four infrastructure providers billing in parallel.
  • Competitor watch. The first run found 14 of 55 competitors active in the week. The competitor list is for his separate business, Rank Spark AI (AI-search visibility).
  • Meetings. The built-in Fathom connector can fetch summaries, transcripts and action items directly. He first routed it through Composio and then noted that was not needed.

The three-part setup around the tasks (August 2026)

The August video places scheduled tasks as the third of three things he sees in the businesses getting the most from Claude:

  1. Static context files referenced from CLAUDE.md. His rule: if you have typed it into a chat more than twice, it should be a file. The three common files are brand voice, visual identity and business facts (assets, ICP). Details in Brand Context Files.
  2. Skills you do not write from scratch. Use Anthropic’s skill-creator, record the process with Record a skill, or start from existing packs. He points to Anthropic’s small-business plugin (Settings → Plugins → Browse; he counts 31 skills, including invoice chase, month-end close, margin analyzer and a Canva creator) and two free open-source packs by the same author, Claude SEO (32 skills) and Claude Ads (33 skills across 12 platforms; he had not tried it).
  3. Scheduled tasks that you supervise instead of run. A daily SEO article task that calls the Claude SEO skill is his example of a 15-minute setup. Runs send a notification to the mobile app.

Try It

  1. Pick one weekly admin job that ends in a checkable state (every row matched, every competitor checked).
  2. Write the done criteria as one sentence before anything else.
  3. In the desktop app: Schedule → New task → Create with Claude. Paste the job and the done criteria and let it interview you.
  4. Confirm each app it needs is reachable as a connector. If an app has no connector, or you need several accounts of one app, add Composio as a custom connector, or a remote MCP server you host.
  5. Accept the trial run on past data and read its output before the first scheduled run.
  6. Start with read-only tasks (audits, reports) before tasks that write to your accounting system.

Implementation

Tool/Service: Claude desktop app scheduled tasks running in Anthropic’s cloud; connectors (built-in, custom remote MCP, Composio). Setup: Schedule tab → New task → Create with Claude, or write the name, description and frequency by hand. Custom connectors: Settings → Connectors → Add custom connector → remote MCP server URL. Cost: Included in paid plans per the creator. Composio free plan: 20,000 requests a month (his figure in July; see Team Shared Memory for the August pricing he quotes). Integration notes: In July 2026 cloud tasks could not read local folders or hold API keys. Since then, scheduled tasks can live inside cloud projects that hold uploaded context files (Claude Cowork), which may remove the need for the Notion workaround.

Open Questions

  • Anthropic’s loop-engineering post he cites is not linked in the transcript; it was not located for this article.
  • Whether cloud tasks have since gained a credential store or local-file access beyond project uploads was not checked.
  • The Claude SEO and Claude Ads packs are named by caption only (“Daniel” is the author); the repository URLs and licences were not checked.
  • His usage figures (17 receipts, 15% rise, 14 of 55 competitors) are from his own account and were shown on screen.