Recently, I noticed that Claude's web UI had a switcher: "Chat" vs "Cowork."
Curious to know when I should use "Chat" vs "Cowork," I asked Claude (Chat) what the difference was, and it responded with:
Chat is a conversation, Cowork is a hand-off. Chat is a conversation you steer turn by turn. Cowork is a delegation: you describe the goal, Claude works across your files and tools, and you come back to a finished result
What I found interesting was "across your files." Cowork has access to a filesystem and a sandbox. I confirmed it, when I asked Cowork to run ls.
Without me noticing, Anthropic pivoted all of their agents to computer use agents with a full sandbox. Cowork is just like Claude Code. They both run in sandboxed environments with full computer use.
Is this just a Claude trend? Is everyone else doing this too?
Sandboxes. Sandboxes everywhere
Cowork as a computer-use agent led me to look at ChatGPT, Gemini, and Perplexity. I was a bit surprised to see that ChatGPT - despite having Codex and strong computer-use models - didn't have a web-based computer-use agent.
Gemini also didn't have a web-based computer-use agent. Whereas Perplexity has an entire tab on their sidebar dedicated to Perplexity Computer.
The Perplexity Computer landing page is worth a visit. It's quite beautiful.
I keep specifying "web-based" because OpenAI's Codex and Google's Gemini CLI do have computer-use tools like bash and code execution when you run them locally on your computer. But they do not have sandboxed environment that fully run in the cloud. Remote sandboxes let you run long-running tasks without depending on your computer staying on or working with your computer's local files.
Nature vs nurture: environment matters
ChatGPT and Google's AI work perfectly well. I notice that my wife's go-to is the Google AI Overview, and she uses ChatGPT to plan our vacations.
Computer-use may not be needed for search use cases, but when the use case is automating work a user would do on the computer, it makes perfect sense to give an agent the same environment a user would work in.
Yes, MCP connectors give ChatGPT or Gemini the tools to interact with 3rd-party apps on your behalf, but the reasons I believe Claude is all-in on sandboxes on top of MCP, skills, and templates are:
- The flexibility of bash
- The reliability hierarchy
Bash is one tool, that's every tool
With just a tool to run bash, agents can read files, write files, run code, and call APIs. If
that's not enough, you can also install and run other CLI tools like git, vercel, and docker.
There's a reason why every product started building MCPs and CLIs - agents retrieve context on how to run CLI commands
with the help command, then run commands all with bash.
jackmu@Mac % fl0w -h
fl0w — agent harness with MCP, skills, and workflows
Usage:
fl0w Start the interactive TUI
fl0w help | -h | --help Show this help
fl0w login Sign in to fl0w (model calls use your fl0w account)
fl0w logout Sign out
fl0w whoami Show signed-in account
fl0w sync Sync config, auth, skills, and tools with fl0w-work; push pending session/workflow writes
fl0w reauth [server...] Connect to all (or the named) MCP servers, running OAuth where needed
fl0w mcp [list|add|remove] Manage MCP servers in mcp.json (fl0w mcp --help)
fl0w secrets [list|add|remove] Manage credentials the agent can use without seeing them (fl0w secrets --help)
jackmu@Mac % fl0w sync
Bash is not just for CLI tools. Bash commands are a way for agents to read, write, and update files in a linux sandbox. Agents can use files to load, unload, and manage their own context on their own, circumventing the need for external databases like vector databases, prompt libraries, and session history in some use cases.
With deep training of bash command use, LLMs are equipped to run bash commands, debug bash errors, to complete a variety of tasks - from coding tasks to data analysis to presentation.
Reliability is a ladder
For agents like Claude Cowork, computer-use doesn't stop at bash. Cowork has browser-use and can even use computer vision to run the computer directly to truly give agents an environment that emulates a user's work computer.
If agents can use browsers and the computer desktop itself, why do agents need MCP, CLI, or bash? It's because each interface varies in reliability, where MCP tools are the most reliable and computer vision the least.
Most reliable
│
1. MCP tool call Slack MCP → call message tool directly
│
2. Bash commands & script CLI command or a Python script that calls the Slack API
│
3. Browser use (CDP) Chrome DevTools Protocol: read HTML, click
│ forms, navigate
▼
4. Screenshot + computer vision over desktop apps
computer vision slowest, most expensive, least reliable
│
Least reliable
When agents are given a task, they take the steps most likely to succeed and where the context is the least muddied.
- MCP gives agents clean context with tool descriptions and clear inputs
- Bash commands and scripts can likewise run well-documented CLI commands and call varying levels of documented APIs
- Browser use drives a Chrome browser through a protocol called CDP - web pages can come with a lot of noise like different authentication methods, iframes, and javascript, which is why browser use also uses screenshots when CDP is not enough
- Computer vision is the most general and emulates human users the most - take a screenshot of the desktop, decide the next action
You can see that MCP and bash are purely text-based, where reasoning models can cleanly ingest and generate text outputs. On the other hand, browser-use and computer vision driven tasks depend on image modesl which are slower, more expensive, and less reliable with our current models.
Fl0w was built for agent-driven automations, so you can try MCP, bash, and browser use automations with the fl0w CLI. Computer vision is not quite there in my opinion.
Agents like Cowork, Perplexity computer, and fl0w run in their own computer sandboxes, not to use computer vision and browser-use all of the time, but to have different options for completing a task with the highest probability. Whereas MCP-based agents like ChatGPT are reliable and great for direct search and API use, agents with computer use can default to MCPs and APIs, but fallback on browser use and computer vision when needed.
Anthropic at the forefront
Claude's preference toward sandboxed computers are a headwind that computer-use will be the default. Computers were made for human use, and through millions of neural network training sets, LLMs are now also a user for computers.
Where we go next is computers for humans and agents to work together in. It's already happening with local agents like Cowork and Claude Code on your laptop. Why not have a personal cloud computer on fl0w.work for agents to work when your laptop is sleeping?
Try it free at fl0w.work.