Lesson 44 of 1570
Tools an Agent Might Have: Filesystem, Browser, Code
Agents are only as useful as their tools. Tour the big three — filesystem, browser, code execution — plus the emerging MCP ecosystem, with examples of what each unlocks.
Lesson map
What this lesson covers
Learning path
The main moves in order
- 1Tools are superpowers
- 2agent tools
- 3filesystem
- 4browser automation
Concept cluster
Terms to connect while reading
Section 1
Tools are superpowers
A language model without tools can only talk. Add a browser and it can research. Add a filesystem and it can build. Add code execution and it can compute. Tools turn 'smart autocomplete' into 'team member who can actually do things.'
The core three
Compare the options
| Tool | What it unlocks | 2026 reference products |
|---|---|---|
| Filesystem | Read, write, move, delete files. | Claude Code, Devin, OpenClaw. |
| Browser | Open pages, click, fill forms, scrape. | Computer Use, Browser Use, Operator, Atlas. |
| Code execution | Run scripts in Python, JS, shell. | Code Interpreter (ChatGPT), Claude Analysis tool, sandboxed shells. |
Filesystem — the quiet workhorse
Once an agent can read and write files, half of knowledge work becomes automatable: summarize this folder of PDFs, rename these photos by date, collect all invoices from 2025, refactor this codebase. Filesystem is the single tool that unlocks the most 'I wish I didn't have to do this' tasks.
Browser — messy but essential
Browser tools let agents use any web app. No API required. The tradeoff: clicking pixels is slow, brittle, and vulnerable to UI changes. OSWorld scores jumped from under 15% in 2024 to 72.5% by early 2026 after Anthropic's Vercept acquisition improved Claude's vision-based perception. Still, browsers remain the hardest domain to make robust.
Code execution — compute on demand
When an agent needs to do real math, parse a CSV, transform data, or generate a plot, it writes code and runs it. ChatGPT's Code Interpreter, Claude's Analysis tool, and Devin's sandboxed environment all work this way. Much more reliable than asking the model to 'calculate in its head'.
The fourth category: MCP tools
MCP (Model Context Protocol) is the open standard launched by Anthropic and now backed by OpenAI and Google. It lets any AI client connect to any tool server with one protocol. As of April 2026, there are over 1,200 community MCP servers — think of them as apps for your agent.
- GitHub MCP — create issues, review PRs, manage branches.
- Notion MCP — read and write your workspace.
- Supabase / Neon MCP — talk to your database.
- Stripe MCP — query payments, create products.
- Zapier MCP — reach 8,000+ connected apps.
- Slack MCP — read channels, post messages.
- Linear, Figma, Google Drive, Calendar — all officially supported.
Next we'll look at where those tools live — on someone else's cloud or on your own machine — and why the difference matters a lot.
Key terms in this lesson
End-of-lesson quiz
Check what stuck
15 questions · Score saves to your progress.
Tutor
Curious about “Tools an Agent Might Have: Filesystem, Browser, Code”?
Ask anything about this lesson. I’ll answer using just what you’re reading — short, friendly, grounded.
Progress saved locally in this browser. Sign in to sync across devices.
Related lessons
Keep going
Builders · 32 min
Your First Agent: A Walkthrough of What It Does
Follow a real agent run step by step — from prompt to result — and see exactly what happens inside. No code yet, just the anatomy of a successful task.
Builders · 32 min
Ollama Basics: Running a Model Yourself
Ollama turns 'I want to run an LLM locally' into a one-line install and a two-word command. Here's the stack, the key commands, and the models worth pulling first.
Creators · 48 min
Computer Use API: Letting AI Click Through GUIs
Computer Use lets Claude see your screen and use it — mouse, keyboard, apps. The capability is real, the gotchas are real. A hands-on look at what works in 2026.