In the previous article I sorted 2,716 posts about "AI employees" and "AI companies" into five stages: role prompt, instruction files, scheduled job, always-on single agent, and multi-agent. At each stage the human's job changes, from asking every time to only approving.
The next question I got was practical: "Then which app should I use, and what can it actually do?" This article answers that for Claude Code, ChatGPT (including dots), other always-on agents, Paperclip, OpenClaw, and running any of them on a Mac mini. For each one I separate what it can do, what it cannot do, who it suits, and who it does not suit.
Product names are a poor guide here. The same product lands in a different stage depending on which feature you use. Claude Code with only a CLAUDE.md file is an instruction-file setup. With Routines it becomes a scheduled job. With Channels it becomes an always-on agent you can message. So this article looks things up in the order what you want to happen → the feature that does it → the product.
Facts here are as of October 14, 2026. Several of them changed more than once this year, so I give dates where it matters.
This article stays close to the tools themselves. It is for readers who want the facts laid out and will decide for themselves which tool fits. The "Which to pick, by what you want" section near the end is also a choice between tools. It does not yet speak to readers who think, "I want to automate my own work, but I don't know what I need." That is the next article. There I will start from everyday work: sorting, summarizing and answering email automatically, automating the bookkeeping, checking Instagram on a schedule, replying to LINE messages automatically. For each kind of work I will explain which apps and which machines you need, in terms that do not assume technical knowledge. Whether buying a Mac mini pays off in your case goes there too. The Mac mini section in this article gives only my own judgment and the general conditions.
The short version
- Sort by two things, not by product name: what starts the work, and where it runs. "What starts the work" is the stage from the previous article (a person asks / a schedule / a message / another agent). "Where it runs" decides whether it keeps working when you close your laptop.
- Claude Code covers four stages depending on the feature. Instruction files on your own machine, Routines in the cloud (schedule as short as once an hour), Channels and Remote Control for messaging a running session, and Projects for parallel threads.
- ChatGPT has two different things. Scheduled tasks, Work and Codex are scheduled jobs. dots is an always-on single agent that runs in OpenAI's cloud and comes back to you on its own. dots needs Pro, Business Premium or Enterprise.
- OpenClaw is free software you run on your own machine and talk to from chat apps. I installed it in a test container. With the default settings, it ran a shell command I asked for without asking for approval. After switching to a stricter preset, it showed an approval card instead. A job set to every minute ran three times in a row.
- Paperclip does not think by itself. It ties Claude Code, Codex, OpenClaw and others together as "workers" with an org chart, tasks, budgets and approvals. That is the multi-agent stage. Its budget does not stop subscription overuse, because subscription runs are recorded at 0 cents.
- A Mac mini is a place, not a tool. It only adds something when you need things that exist only on a Mac (iMessage, macOS apps) or a machine at home (a home IP address), or when you run a local LLM (a large language model on your own hardware). If you use Claude Code or Codex with cloud models, my judgment is that an always-on Linux CPU server is better value.
Body (about 15 minutes)
Two axes: what starts the work, and where it runs
Here are the five stages from the previous article again, with what starts the work at each stage.
| Stage | What starts the work | What the human does (most common) |
|---|---|---|
| Role prompt | A person asks each time | Asks each time (87.1%) |
| Instruction files | A person starts it; work is split inside | Starts it (68.8%) |
| Scheduled job | A time, or an event such as new mail | Reads the result (84.1%) |
| Always-on single agent | You message it, or it resumes on its own | Messages it (61.6%) |
| Multi-agent | One worker hands work to another | Only approves (54.3%) |
The percentages are the most common human role within each type in the previous article's 2,716 posts.
The second axis is where it runs. There are three places: the provider's cloud, your own machine (a laptop, a Mac mini, a rented server), and your own browser through an extension. This decides two things: whether it keeps working when your laptop is closed, and whether it can reach your local files and logged-in sessions.
The chart below places each product's features on the five stages. The point is that one product spans several stages.
And this one shows where each feature runs, and the two lines that matter: whether it needs your computer to be on, and whether it can reach your local logins.
Comparison table
Terms in the table: a heartbeat is a timer that wakes the agent to check whether there is anything to do. cron runs jobs at set times or intervals. A webhook is a trigger fired when another service calls in.
| Tool (feature) | Stage | Where it runs | What starts it | Approval | Cost | Works with laptop closed |
|---|---|---|---|---|---|---|
| Claude Code (CLI / Desktop) | Instruction files to always-on | Your machine | You ask; /loop and Desktop scheduled tasks (from 1 minute); Channels; Remote Control | Six permission modes; interactive default is auto | Subscription limits (Pro $20/month and up) | No |
| Claude Code Routines / Projects | Scheduled job / multi-agent | Cloud | Schedule (from 1 hour), API, GitHub events | Routines mostly run without approval (publishing a new page asks first); Projects asks inside the thread | Same subscription | Yes |
| Claude (formerly Cowork) | Scheduled job | Cloud (all new Pro/Max tasks since 2026-10-06) | You ask; schedules such as hourly, daily, weekly (shorter than hourly not confirmed) | Manual / Auto in the new experience | Same subscription | Scheduled tasks: yes |
| ChatGPT dots | Always-on single agent | Cloud, plus one optional local PC | Conversation, resumes on its own, schedules made in chat, app events | An automatic reviewer decides proceed / ask / leave to you | Pro, Business Premium, Enterprise | Yes |
| ChatGPT scheduled tasks / Work / Codex | Scheduled job | Cloud (web), local (desktop app) | Time; Gmail, Slack, GitHub events (web) | Confirmation before important actions | Plus $20 and up | Web: yes |
| OpenClaw | Always-on single agent (multi-agent possible) | Your machine | 33 chat apps, heartbeat (30 min default), cron, triggers, webhooks | Default runs commands without approval | Free (MIT) + model + machine | Yes, on a server |
| Paperclip | Multi-agent | Your machine (server + database) | heartbeat, routines (cron / webhook / API), task assignment | Hiring and strategy approvals built in; business approvals you design | Free + the "brain" + machine | Yes, while the machine is up |
What each tool can and cannot do
Claude Code on your own machine
- Can: Work with local files and commands. Keep procedures in files (CLAUDE.md, skills, subagents) and force checks to run with hooks. Through Channels you can message a running local session from Telegram, Discord or iMessage. Remote Control lets you continue a local session from your phone.
-
Cannot: Keep running when the computer sleeps. A /loop disappears after 7 days. Chrome control stops at login screens and CAPTCHAs. Computer use does not work in unattended
-pmode. - Suits: Developers who want the AI to work inside their own environment and repositories.
- Does not suit: People who cannot keep a computer running, or who do not want to touch a terminal.
This is how I use Claude Code every day: instruction files and skills, with hooks that block external sends. In the previous article's terms that is the instruction-file stage.
Claude Code in the cloud: Routines and Projects
- Can: Run on a schedule with your laptop closed (once an hour at the shortest), or start from an API call or a GitHub pull request. Projects lets you send several pieces of work to parallel threads and leave them.
- Cannot: Run more often than hourly, or log in to websites. Routines mostly do not ask for approval mid-run (publishing a new claude.ai page is an exception that asks first). Projects asks inside the thread when it needs approval, and you answer there. A green mark in the run list does not prove the work succeeded. Projects is a public beta for Pro and Max; it is not available on Team or Enterprise yet, or from the CLI.
- Suits: Developers whose work already lives on GitHub.
- Does not suit: Work where a person must check before anything is sent, or web work that needs a login.
Claude (formerly Cowork), Claude in Chrome, Claude Tag
Anthropic announced on September 16, 2026 that Cowork is being merged into Claude; it is rolling out to Pro and Max first, with Team and Free later. From October 6, all new tasks on Pro and Max run in the cloud.
- Can: Run documents, research and file tidying on a schedule without writing code, even while your PC sleeps. The Chrome extension handles routine web work that needs your login. Claude Tag shares an @claude in Slack with the team (Team and Enterprise only).
- Cannot: The Chrome extension never buys things, creates accounts or permanently deletes in any mode. Claude Tag is not available on Pro or Max. I could not confirm a schedule shorter than hourly.
- Suits: Business owners and freelancers who do not write code and want routine work done.
- Does not suit: People who want local files handled unattended, or unattended browser work at night.
ChatGPT dots
I tested dots with my real work accounts in the first article of this series. dots runs on a dedicated computer and browser in OpenAI's cloud, and you can connect one local PC.
- Can: Keep working with your PC closed, and come back to you when there is a result or a decision. Talk with the same memory from Slack, Teams or a phone call (you call it). Keep logins saved in its cloud browser.
- Cannot: It is not on Plus, and not on Pro in the EEA, the UK or Switzerland. It cannot call or text you yet. The official docs do not state a shortest interval or a run limit. Actions through your local PC look like your own actions to the other side.
- Suits: People on Pro or Business Premium who want a secretary for mail, calendar and Slack, including non-developers.
- Does not suit: People on Plus, work that must run at an exact time, and anyone who needs AI actions recorded separately from their own.
One point I will not state as settled. OpenAI's docs say existing app permissions also apply to dots. In my test, dots deleted a calendar event without asking once. That is one case against the docs, so treat it as unresolved.
ChatGPT scheduled tasks, Work and Codex
- Can: Run a saved instruction at a set time or when mail arrives, and finish documents, tables and regular reports. The desktop app runs unattended on local repositories. Codex Cloud goes as far as opening a pull request.
-
Cannot: Slack DMs and reactions cannot start a task. The official docs do not describe retries or failure notices. The Codex CLI itself has no scheduler; you call
codex execfrom launchd or cron. - Suits: Plus users who want regular reports, and developers who already use Codex.
- Does not suit: A standalone task that must run more often than hourly. News reports say standalone scheduled tasks run at most once an hour, while OpenAI's learning pages say recurring tasks inside a chat can run every few minutes. I could not open the Help Center page to settle this.
GPT-5.5 was retired from ChatGPT, Work and Codex on October 14, 2026, so older examples that name it are out of date.
OpenClaw (I installed it to check)
OpenClaw is free, open-source software (MIT license) that you install on your own computer or server. You talk to it from chat apps you already use, such as Telegram, Slack, WhatsApp, iMessage or LINE (33 channels in the official list), and it does the work on that machine. It has no model of its own; you connect Claude, GPT or another model. It was called Clawdbot and then Moltbot before January 30, 2026, and is now run by the OpenClaw Foundation.
I installed version 2026.10.3 in a Docker container on my Mac. I connected no business accounts and no chat apps, only the built-in web screen (the Control UI). For the model I signed in with my ChatGPT subscription through OpenAI's device-code login, so no extra fees.
A chat round trip worked. I asked what it can do and what it needs my approval for. It answered that it can edit files, run commands and browse, and that it asks before destructive or external actions.
With the default settings, it ran a shell command without asking. I asked it to run uname -a && whoami && date -u, and it ran and showed the output. The command-approval summary for the default setup says "no approval prompts."
With the "cautious" preset, it showed an approval card instead. I switched to exec-policy preset cautious (allowlist, ask on miss, deny on fallback). When I asked it to save a note to memory, it wanted to read a file and showed an approval card with "Allow once," "Always allow" and "Deny." One more thing happened along the way: after the switch, commands tried to run inside an isolated environment (bwrap), which failed to start, and the request to run outside it was refused, so even a plain ls did not run. Because the model was connected through Codex on my ChatGPT subscription, the approval card was titled "Codex app-server command approval." I did not separate whether this isolation and approval came from OpenClaw's settings or from the connected Codex.
A job set to every minute ran three times in a row. I added cron add --every 1m with "reply with the current UTC time." It ran at 09:08, 09:09 and 09:10 UTC, and each time the model answered with the right time. Every run was marked as an error only because I had connected no chat app to deliver the result to. After the third run, the next run was scheduled five minutes later (09:15 UTC); I did not check why. On the same screen you can see what a fresh install already has scheduled: a heartbeat every 30 minutes, a daily memory job at 3:00 UTC ("Memory Dreaming Promotion"), and a weekly skill review.
Memory was written to a Markdown file in its workspace. After switching to the no-approval preset (exec-policy preset yolo), I asked it to remember that my test project is called AICO-003 and that I prefer short answers. OpenClaw added those two lines to a file called USER.md and told me which file it wrote. It is plain text, so a person can open it, read it and fix it.
- Can: Live on your machine and take requests from 33 chat apps. Run on schedules down to a minute, and on condition triggers checked as often as every 30 seconds. Use your own logged-in Chrome, and run shell commands. Pick and mix models. Memory is plain Markdown files in its workspace (USER.md, MEMORY.md, daily notes).
- Cannot: Be shared safely by people who do not trust each other (the docs say it is not designed for that). The defaults are not safe for real accounts: commands run without approval and the sandbox is off. The official FAQ says small local models are usually not practical.
- Suits: People who can manage their own computer or server, want to make requests from chat, or need jobs every few minutes.
- Does not suit: People who do not want to own exposure settings, permissions, skill review and updates themselves (722 security advisories published so far, releases almost every day), or who want to connect an important inbox on day one.
I do not read the 722 advisories only as a danger sign. They are also a record of problems found and fixed in the open. What I can state as fact is the setting: by default the command mode is full (runs without asking), and the docs recommend auto, where an AI reviewer checks anything not on the allowlist and passes risky ones to you.
Can you use a Claude subscription with it? As of October 14, 2026, Anthropic's support article (updated October 7) says third-party apps can use your subscription limits. Anthropic's terms also say Pro and Max limits assume ordinary individual use. The rules changed five times in 2026, and OpenClaw itself warns that Anthropic can change this again. For OpenAI, OpenClaw's docs say subscription sign-in is allowed, which is what I used.
Other always-on agents
- Run on your own machine: Hermes Agent and NanoClaw are similar to OpenClaw. NanoClaw isolates work in Docker.
- Run in the provider's cloud: Gemini Spark, Manus, Microsoft Copilot Cowork, Lindy and Zapier Agents start from a schedule or from events such as new mail, and need no computer of yours.
- Cannot: Gemini Spark does not work with work or school accounts. Copilot Cowork needs a Microsoft 365 Copilot license. Manus credit use is hard to predict in advance.
- Suits: If your work is mostly Google, Gemini Spark; Microsoft 365, Copilot Cowork; many connected apps, Zapier.
- Does not suit: Handing over whole jobs with many judgment calls (Zapier is best at connecting apps).
What changes when you add Paperclip
Paperclip is open-source software that gives AI agents an org chart, assigned work, budgets and approvals. It does not think by itself. The workers inside are Claude Code, Codex, OpenClaw and others. Adding it moves you to the multi-agent stage: a worker finishes a task and hands the next piece to another worker without going through you.
I run it as "AI Office," my own secretary. As of October 14 it has 7 workers, all running Codex on my subscription, 4 routines (for example the mail worker wakes at minute 7 of every hour), issues numbered up to 668, and 72 approvals decided on screen.
- Can: Keep a record of who did what, prevent two workers from taking the same task, give each worker different permissions, and run cron and webhook routines with an owner and a history.
- Cannot: Think without a separate "brain." Stop subscription overuse with its budget: runs on a subscription are recorded at 0 cents, so the budget never trips. I put a separate guard in front of Codex that refuses to start it once plan usage passes 95%. Memory is still in progress on the roadmap, and the hosted version is still a waitlist.
- Suits: People who want different agents under one organization, recurring work with an owner and history, and who can look after an always-on server themselves.
- Does not suit: People who only want one agent to do one recurring job (Routines or ChatGPT scheduled tasks are enough), or who cannot dig into a database and logs when it stops. Between October 7 and 9 mine hung and had to be restarted, and lost its database connection, which I spent three hours tracing.
How OpenClaw differs from Claude Code, dots and Paperclip
- Versus Claude Code: Claude Code is something you start and work with. Channels lets you message a running session, but the session lives on your computer. OpenClaw is built to stay running and to be reached from chat apps, with its own heartbeat and scheduler.
- Versus dots: Both are always-on single agents. dots runs in OpenAI's cloud, with OpenAI's reviewer deciding when to ask you. OpenClaw runs on your machine, and you set the approval rules yourself. The default is no approval.
- Versus Paperclip: OpenClaw is one worker (it can route to several agents). Paperclip is the organization around workers, and OpenClaw can be one of them.
What changes on a Mac mini
A Mac mini is not another tool. It is one value of "where it runs": your own machine, set not to sleep. Apple's own page this August called it ideal for always-on AI agents, and OpenClaw's FAQ says a Mac mini is not required; a VPS (a rented virtual server) or a Raspberry Pi also works. These are not in conflict. It depends on the job.
- What it adds: It does not sleep, and it comes back after a power cut (2024 and later models on macOS Tahoe 26.5 or later). It can use iMessage, macOS apps, a home IP address and logged-in browsers at night. It can run local LLMs (up to 64GB of memory on the M5 Pro). It separates the AI from the Mac you use every day.
- What it does not change: Cloud products keep the same limits. One machine is not redundancy. With FileVault on, coming back unattended after a restart is hard.
My own decision: I own a Mac mini, but I do not use it for AI. If you have no other machine for a local LLM, you need a Mac mini. But as long as I use Claude Code and Codex with cloud models, machines like the always-on Linux CPU servers my company runs (two machines we call ci1 and ci2) are better value. For LLMs I bought an NVIDIA DGX, so I do not see a need for a Mac mini even for that.
From what I researched, a Mac mini makes sense when you have work that only a Mac can do, such as iMessage or macOS apps, or when you have no other machine to run a local LLM.
Which to pick, by what you want
| What you want | Start with |
|---|---|
| The AI works inside my code and files while I watch | Claude Code (or the Codex CLI) |
| A report or summary at a set time, laptop closed | Claude Code Routines, Claude scheduled tasks, ChatGPT scheduled tasks |
| A secretary I message, who comes back to me | ChatGPT dots (Pro), or OpenClaw if I can run a server |
| Jobs every few minutes on my own machine | OpenClaw, or launchd / cron calling Claude Code or Codex |
| Several agents passing work to each other with approvals | Paperclip, with Claude Code / Codex / OpenClaw as workers |
| Mac-only work at night (iMessage, macOS apps) | Any of the "your machine" tools, on a Mac mini |
Points I am not stating as settled
- Whether a Claude subscription can run OpenClaw. Yes as of October 14, 2026, but it changed five times this year.
- The shortest interval for ChatGPT scheduled tasks. Reports say hourly; OpenAI's learning pages say in-chat recurring tasks can be minutes.
- Whether dots respects app permissions. The docs say yes; my one test showed a deletion without asking.
I tested OpenClaw only in a container with no accounts connected. I have not tested Routines, Projects, Channels, Claude Tag, Hermes or a Mac mini setup myself; those parts are from public documentation.
Top comments (1)
The zero-dollar accounting problem is where multi-agent orchestrators quietly blow up. When tasks register at zero marginal cost because of a flat subscription, the coordinator has zero incentive to throttle dispatch. A circuit breaker in front of the execution runner is the only thing that actually protects the rate limit.