Giving an agent hands: MCP servers and browser control
The Model Context Protocol (MCP) is the standard way to connect an AI agent to external tools and data, and the modelcontextprotocol/servers repository is the best reference implementation to learn from. For systems that expose no API at all, browser-use drives a real browser as a fallback. Both are free, open source, and the difference between a chatbot and an agent that does something.
Free. 2 repositories, no paywall. Tip only if it helps.
A model with no tools is a chatbot. The whole value of an agent is that it goes and does a thing: reads the calendar, files the ticket, pulls the order, updates the record.
So the tool layer isn't a detail at the end of the build. On our projects it's usually where most of the engineering actually goes, which surprises people who budgeted for prompt writing.
What changed recently is that this stopped being bespoke. Before MCP, every integration was a hand-rolled function definition, coupled to one framework and one model, rewritten whenever either changed. MCP made the connection a protocol, so a server you write once is reachable from anything that speaks it. Read two of the reference servers and you'll understand it better than from any explainer, including this one.
Then there's the other half of reality. The system your client depends on has no API, the vendor has no roadmap for one, and nobody is switching. That's where browser-use comes in: an agent driving a real browser. It works, we've shipped it, and I want to be straight about it. It is brittle by nature. A layout change breaks it. Budget for maintenance from day one and never put it anywhere silent failure costs someone money.
One security note, because it gets skipped. The moment your agent reads content from the open web or from a document a stranger uploaded, that content can contain instructions. Treat everything a tool returns as data, never as orders, and keep destructive actions behind a confirmation a human gives.
This week: list every tool your agent can call, and mark which ones can write. If that second list is longer than you expected, that's the finding.
Tools: the part that makes an agent useful
A model with no hands is a chatbot. This is how you give it hands, and how you keep the connection standard instead of bespoke.
- modelcontextprotocol/servers90,335· TypeScript · MIT
Reference MCP servers. Read two of them and you will understand the protocol better than any explainer, this one included.
- browser-use/browser-use114,628· Python · MIT
When the system you need has no API and never will. Brittle by nature, so budget for it breaking.
Star counts are a snapshot taken 2026-09-14, not live figures. Licences are as published by each project, so check them yourself before you ship commercially.
Want the whole list in your inbox?
Every repository above, with the notes, in one email you can search later, plus a note from me whenever we publish something new on AI and where mobile development is going. That's the only trade.
Tip if it helped. Only if it helped.
The library is free and stays free. You already have it, whether or not you click anything here. If it saved you an afternoon, you can name your own amount. It goes toward the time we spend testing these so you don't have to.
You can change the amount on the next screen. Secure checkout by Stripe, and we never see your card.
Questions people actually ask
Ready to Build Your Application?
Let's discuss your project. Our team will help you build solutions that transform your business.