Claude, ChatGPT and Codex all let you save a prompt and run it later without sitting there, but they differ on where it runs, what can start it and what happens when it needs permission. Check four things for each product: location, trigger, access and controls.
Why compare them at all
Teams rarely use one assistant. A developer may run Claude Code in a terminal, a product manager may use ChatGPT on the web, and a platform engineer may use Codex in an editor. Each has its own idea of a scheduled task. If you write a governance policy for only one, the others quietly become exceptions. This post lays out what the vendors' documentation says, read on 19 October 2026, so you can see the shared questions and the real differences. It does not rank the products or say which to buy.
The comparison
| Question | Claude Code and Cowork | ChatGPT and Codex tasks |
|---|---|---|
| Where can a task run? | Cloud routines on Anthropic managed infrastructure, local Desktop tasks, session loops in the terminal, and remote Cowork tasks | Web and mobile tasks that use uploaded files and connected tools, and desktop tasks that can work with local projects |
| What can start it? | Schedule, API call and GitHub event on routines. Schedule only for local tasks and loops | Time schedules, and on eligible plans events from Gmail, Slack or GitHub on web and mobile |
| Can triggers combine? | Yes, one routine can have several triggers | A task cannot combine event triggers with a time based schedule |
| Local files? | Desktop tasks and loops, yes. Cloud routines, no, they use a fresh clone | Desktop tasks, yes. Web tasks cannot use a folder on your computer |
| Permission handling | Cloud routines have no permission picker and run without approval prompts. Desktop tasks have a permission mode per task | Unattended tasks run under your default sandbox settings. Read-only and workspace-write fail when they need more access |
| Admin controls | Routines toggle for Team and Enterprise owners, Cowork admin toggle, an environment variable to disable the cron scheduler | A permission for event triggered scheduled tasks, plus admin requirements for sandbox and approval limits |
| When your computer is off | Cloud routines and remote Cowork tasks keep running. Desktop tasks and loops do not | Local file tasks need the computer on and the app running |
Three shared patterns
A saved prompt acts as the assignment. In both families the prompt is written ahead of time and then runs with whatever access the task was given. Anthropic's routines page states that the fired prompt is not live user input and cannot serve as approval for actions during the run.
Unattended means fail closed or fail open. OpenAI describes sandboxed runs where calls that need more access fail. Anthropic describes cloud routines that run shell commands and connectors without stopping. Neither is wrong, but you need to know which you have.
The creator's identity is usually the actor. Anthropic states that routine actions through your connected GitHub and connectors appear as you. OpenAI's page notes that connected apps must be authorized before creating a task, and says little about how credentials are stored.
What the differences mean for governance
- Inventory by product. A single register with a product column covers all of them, but discovery steps differ. See inventorying scheduled tasks across assistants.
- Decide event trigger policy separately. Event triggers let outsiders start runs, which changes the threat model. See event triggered or time based tasks.
- Set a rule for local versus cloud. Local tasks depend on a person's machine and files. Cloud tasks depend on connectors and environments.
- Check model retirement. OpenAI's page notes tasks that use retired models need updating and names a retirement date in October 2026 for one model in Codex. Any task pinned to a model has a lifetime.
- Keep a single kill switch list. For each product, write down where to pause, revoke and disable.
Worked example: one policy, three products
A team writes one rule: scheduled tasks that write to shared systems need a sponsor and a registered connector list. To apply it, the owner of a Claude routine lists repositories and connectors from the routine page. The owner of a ChatGPT task lists the connected apps and the sandbox mode. The owner of a Codex automation lists the project, the sandbox level and whether it runs in a worktree. The register has the same columns for all three, filled from different screens. The rule is the same and the evidence differs.
Try it yourself
Click through it at your own pace, or open it full screen.
Watch how an assistant is connected to a tool and what it is allowed to ask for, then compare that with the connector choices in your own scheduled tasks.
Where Safeguard fits
Safeguard offers an MCP Guardrail, agent identity, and policies and guardrails, which apply whichever assistant is calling. The Safeguard CLI can also write code, and the download page lists a Local agent and a Cloud agent as Available.
Quick check
Do all three products run tasks when my laptop is closed?
Not all tasks. Cloud routines and remote Cowork tasks do. Desktop tasks, loops and local file tasks need the computer on and the app running.
Can ChatGPT tasks start from Slack or Gmail?
According to OpenAI's documentation, on ChatGPT web and mobile, eligible plans can start tasks from Gmail, Slack or GitHub events. Event triggers are not available in the desktop app, Codex CLI or IDE extension.
Which product has the safest scheduled tasks?
That is not a question these pages answer. Safety depends on how you scope access, approvals and connectors in each one.
Sources
- Run prompts on a schedule (Claude Code docs)
- Automate work with routines (Claude Code docs)
- Schedule recurring tasks in Claude Code Desktop
- Schedule recurring tasks in Claude Cowork (Help Center)
- Scheduled tasks and automations (OpenAI docs)
- Safeguard download page
Last reviewed: 19 October 2026.
