Boring AI vs Claude Cowork

Cowork and Boring agree on the thesis: describe the outcome and let the model work the steps. Anthropic built Cowork for a person handing off document, data, and research work, and it is very good at that. It reads and writes the folders you connect, coordinates sub-agents on bigger jobs, and hands back spreadsheets with working formulas, decks you can open in PowerPoint, and Claude Docs your team can edit with you. New Boring agents run on Claude Sonnet 5.5 by default, so the two products share an engine. If the work lives in your folders, Cowork is the right tool: Boring agents don't open your folders or drive your browser.

Boring is built for a workflow rather than a person: the weekly invoice chase, the onboarding that touches nine systems, the triage queue. Several people edit those, review what they did, and answer for the result, and the person who should approve a step is often not the person who set the work up. Boring puts agents in a workspace, with roles, private or workspace visibility, live co-editing, and one shared record per run. Every change to a definition is a versioned diff you can roll back. And while Claude is the default, it isn't the only option: managed GPT, Gemini, and Grok need no key, or you can bring your own provider key.

The first version of this page, in July 2026, said every Cowork task ran on your machine and schedules waited for the app to reopen. That was wrong, and we corrected it in August. Today the argument is gone entirely: Cowork runs in Anthropic's cloud by default, scheduled tasks run with no device online, and from 6 October 2026 new Pro and Max tasks run in the cloud. Team and Enterprise keep local sessions, and Dispatch, the desktop-run path from your phone, is closed to new users. When a task needs something on your computer, such as a connected folder, the built-in browser, or computer use, Cowork reaches it through the desktop app while it's open.

So "your laptop is not a server" is no argument against Cowork, and we've stopped making it. Background work and asking before risky steps are now standard for personal agents. The question that's left is who operates the workflow. In Anthropic's docs, a Cowork task starts because a person asked or a schedule fired. A Boring run also starts from a webhook, an inbound email, a submitted form, a SaaS app event, the workspace API, an MCP client, or another agent. When a run fails, you can resume eligible work from the last completed step, and in a numbered step plan a write whose outcome is uncertain is never repeated automatically.

Anthropic's safety guide for Cowork is candid about scheduled tasks, which run while nobody is watching: "Don't schedule tasks that access sensitive files, send messages on your behalf, make purchases, or take other actions that are difficult to undo." It also asks you to review outputs after each run. For a personal agent working alone, that is exactly the right advice.

It also describes the work Boring is built for. A company's recurring work is full of consequential steps: the email to a customer, the refund, the change to a record. An unattended Boring run can reach that step because, by default, it stops there and waits in a shared Tasks inbox. The request leads with the exact tool call and its scrubbed arguments, goes to the reviewer you assign with a due date, reminders, and a configured overdue outcome of reject or skip, and that reviewer can approve, reject, or modify the exact arguments before the call runs; the agent's owner or an admin can reassign it. Exact-action grants decide what an agent can call at all, so a send action can be withheld entirely rather than prompted against.

Cowork's controls are real and configurable. You choose Manually approve, Automatically approve, where Claude reviews each action for safety, or Skip all approvals, and each connector tool can be set to Always allow, Needs approval, or Blocked. Team and Enterprise admins can turn off Automatically approve, and by default members approve write-capable connector tools task by task. Team and Enterprise owners can stream prompts, tool calls, file access, and human approval decisions to their SIEM through OpenTelemetry, and on Enterprise the Compliance API attributes every session to a user.

Whenever a person approves a step, it is the person whose session it is, and sessions can't be shared. That gives a security team a stream of each person's events. Boring gives the team one reviewable run, with the approval attached, that a colleague can open. Each gate waits in a shared Tasks inbox, where the assigned reviewer, the agent's owner, or a workspace owner or admin decides. Tool payloads are kept 90 days; the outcome, output, and step list stay. The workspace assistant (⌘J) shows which agents need attention and what's waiting on a person across every agent you can access, and changes nothing until you confirm.