Guide
What Claude Cowork Is, and How to Watch an AI Agent While It Works
Cowork is a real Anthropic product, not a generic term. Here is what it actually does, and how developers watch an AI agent while it works on their machine.
Claude Cowork is a real product: Anthropic's research-preview desktop agent, launched on January 12, 2026, built for people who are not developers. You give it access to one folder, describe a task in plain language, and it works through a visible to-do list, turning a folder of receipt photos into a spreadsheet or drafting a report, then hands you the finished files. It reports progress as a checklist inside the app, not as a CPU or resource meter. If you actually want a live, resource-based pulse on an AI coding agent running in your terminal, that is a different and more technical problem, and the tools for it are Activity Monitor, htop, Docker stats, or a desktop app built for that job, such as Forkbench.
What Claude Cowork actually is
Claude Cowork is a real product. Anthropic launched it on January 12, 2026 as a research preview aimed at people who are not developers. Anthropic has described it as Claude Code for the rest of your work: the same idea of handing a task to an agent and getting a finished result back, built for office tasks instead of a codebase.
You give Cowork access to one folder on your computer, describe what you need in plain language, and it works through the task in the background. It can turn a folder of receipt photos into a spreadsheet, draft a formatted report, rename and sort downloads, or research something on the web and write up the findings. It runs through the Claude Desktop app on macOS and Windows, through claude.ai in a browser, and through the Claude mobile apps, on the Pro, Max, Team and Enterprise plans.
Cowork is built for a different job than a coding agent in a terminal. It edits documents and files inside one folder you grant it. It does not open a shell, run arbitrary commands, or work across your whole filesystem the way a terminal-based coding agent does. That distinction matters for the rest of this page, because the kind of monitoring that makes sense for each one is different.
- Anthropic launched Claude Cowork as a research preview on January 12, 2026.
- It runs through Claude Desktop (macOS and Windows), claude.ai, and the Claude mobile apps, on paid plans.
- It works inside one folder you grant it, not across your whole machine or a terminal.
Why Cowork has no CPU pulse to configure
If you searched for something like cowork agent ai setup pulse hoping to configure a live dashboard, there is nothing to set up. Cowork already shows its progress inside the app as a running to-do list: what it has checked, what it is doing now, what is left. That is a task tracker, not a measurement of processor load.
A CPU or memory pulse only makes sense for something that is a real operating system process on your machine, with its own process ID that a tool like Activity Monitor can inspect. Cowork's work happens through Anthropic's own infrastructure and the Claude Desktop app, so there is no separate local process for you to attach a resource monitor to.
This is a different situation from running a coding agent, such as Claude Code or an open source agent like Cline, directly in your own terminal. Those agents are ordinary local processes, and the rest of this page is about watching that kind of process, which is what most people mean when they ask for a live pulse on an AI agent.
- Cowork's own progress indicator is a to-do list inside the app, not a resource meter.
- There is no local process ID for Cowork that a system monitor could track.
- A terminal-based coding agent is a real local process, which is where a CPU pulse applies.
Watching a local agent process with tools you already have
Every major operating system can already show you what a process is doing. On macOS, open Activity Monitor, search for the agent's process name (often python, node, or the name of the CLI tool), and set the update frequency to Very Often under the View menu for a one-second refresh.
On the command line, htop gives you the same information without leaving the terminal. Run it in its own window and filter for the agent's process. You will see the CPU percentage climb while the agent is actively working and drop while it waits on a response.
If the agent runs inside a Docker container, docker stats shows the same thing for containerized workloads. A formatted command such as docker stats --format "table {{.Container}}\t{{.CPUPerc}}\t{{.MemUsage}}" gives you a readable, live-updating table for every running container, which helps when several agents run at once.
- Activity Monitor on macOS, set to a one-second refresh, works with zero setup.
- htop gives the same view from the terminal, including for remote machines over SSH.
- docker stats shows live CPU and memory for each container when agents run inside Docker.
A closer look inside VS Code or your own script
If your agent runs as a VS Code extension or inside its integrated terminal, VS Code has a built-in Process Explorer. Open the command palette and run Developer: Open Process Explorer to get a tree view of every active process VS Code is managing, each with live CPU and memory numbers.
If you are building your own agent, you can read the same numbers from inside your code with Python's psutil library, which exposes per-process CPU and memory figures you can sample on a timer and print, log, or feed into your own dashboard.
Both approaches give you raw numbers, not meaning. A high CPU reading could mean the agent is making good progress or that it is stuck in a loop. You still have to interpret it against what the agent is supposed to be doing.
- VS Code's Process Explorer command shows live CPU and memory per process and extension.
- Python's psutil library lets you build a custom pulse into your own agent code.
- Raw CPU numbers do not tell you whether the agent is making progress or stuck.
Offline agents need a second pulse: the model server
Running a model completely offline through Ollama or LM Studio changes what you need to watch. The agent logic itself might use very little CPU while the local inference server does the heavy work, so you need to track both processes, not just one.
On macOS, that means watching both your agent's process and the Ollama or LM Studio process at the same time in Activity Monitor. The agent process shows you the decision loop. The model server process shows you the actual token generation.
This matters for your hardware, not just curiosity. Local inference on a laptop produces real heat and drains the battery fast, so a pulse on the model server is also how you decide when to let the machine cool down or plug in before you keep going.
- Offline setups need you to watch the agent process and the model server process separately.
- The model server, not the agent logic, is usually where the CPU and GPU load actually is.
- Use the pulse to manage heat and battery drain during long local inference sessions.
How Forkbench's live pulse works, and what it is not
Forkbench is a desktop app for running coding agents in a real terminal on your Mac. Each terminal tab has a built-in pulse indicator that quickens as the agent works harder, driven by the agent's actual CPU usage and output rate on your machine.
The pulse is not a token meter. It has nothing to do with how much you are spending with a model provider. It is a measurement of local activity, so you get the same signal you would get from htop, but next to the terminal where the agent is actually working, with a red flag and a count that appears when the agent is blocked and needs you.
Forkbench is a separate product from Claude Cowork, and the two are not connected. Cowork is Anthropic's own agent, running through Anthropic's infrastructure, for non-coding tasks inside one folder. Forkbench is a terminal environment for coding agents like Claude Code and Codex, and it has no integration with Cowork or with any other vendor's own hosted agent product.
- The Forkbench pulse tracks real CPU usage and output rate, not API tokens or spend.
- A red flag with a count appears the moment an agent is blocked and needs a human.
- Forkbench is not connected to Claude Cowork or any other vendor's hosted agent product.
Picking the right way to watch your agent
If you run one script occasionally, htop or Activity Monitor is enough. It takes no setup and gives you everything you need to confirm the process is alive and working.
If you run several agents in containers, docker stats isolates each one so you can see which container is actually busy. If you are building your own tool, a custom psutil-based pulse gives you full control, at the cost of writing and maintaining that code yourself.
If you want a pulse that is already built into the place where your coding agents run, without opening a second window, that is the gap a dedicated desktop app like Forkbench fills. It trades some of the granularity of a custom script for a dashboard you do not have to build.
- Use htop or Activity Monitor for a single script with zero configuration.
- Use docker stats to isolate resource use across several containerized agents.
- Use a dedicated desktop app when you want the pulse already built into your terminal.
Related: How Forkbench handles your data, The real problems with managing several coding agents, Answer a blocked agent from your phone, Download Forkbench
Frequently asked
Is Claude Cowork a real Anthropic product?
Yes. Anthropic launched it on January 12, 2026 as a research preview, built for office-style tasks rather than coding, available through Claude Desktop, claude.ai, and the Claude mobile apps on paid plans.
Does Claude Cowork show CPU or resource usage?
No. It shows its progress as a to-do list of steps inside the app. There is no local process ID or resource dashboard to configure, because the work runs through Anthropic's own infrastructure, not a process on your machine.
Can I use Activity Monitor to watch a local model like Ollama?
Yes. Ollama and LM Studio run as ordinary local processes, so Activity Monitor, htop or docker stats will show their CPU and memory use like any other process.
Does the Forkbench pulse track my API or token costs?
No. It is driven by CPU usage and output rate on your machine, not by tokens, and it has no connection to your model provider's billing.