Five Agent Tools Worth Your Weekend

• browser-use/browser-use (GitHub) - Over 107,000 stars for a Python library that lets an AI agent click, scroll and read a website the way you would. Most sites still assume the visitor has eyes and a mouse, not an API key. This is the bridge for that gap. If you have ever tried to get an agent to fill in a form or scrape a page that fights back, this is the tool everyone converges on eventually. Worth running early rather than reinventing your own Selenium wrapper badly.

• upstash/context7 (GitHub) - Context7 solves a specific, boring, constant problem. The model's training data is always out of date and your dependencies are not. It's an MCP server from Upstash that pulls current documentation straight into the model's context, so it stops confidently suggesting a method that got deprecated two versions ago. 60,076 stars says this pain is universal. If you're pairing Claude with a fast-moving framework, this belongs in your setup rather than in a bookmark you meant to try.

• headroomlabs-ai/headroom (GitHub) - Everyone building with LLMs eventually hits the same wall. Tool outputs and RAG chunks are stuffed with tokens the model never needed. Headroom compresses logs, files and tool results before they reach the model, with a reported 20% reduction in tokens. That's not a huge number on its own. Stack it across every call in a production agent and it starts paying for the engineer who installed it.

• microsoft/ai-agents-for-beginners (GitHub) - Eighteen Jupyter notebooks from Microsoft, over 70,000 stars, and genuinely aimed at people who have not built an agent before. It's not Claude-specific and it won't teach you anything advanced. What it does well is give a structured on-ramp to a newcomer, which is rarer than it should be given how much of this space assumes you already know the vocabulary.

• HKUDS/nanobot (GitHub) - A lightweight, self-hosted personal agent framework in Python, tagged under the model context protocol topic and sitting at 46,472 stars. Self-hosted matters here. You keep the data, you keep the config. Nothing routes through someone else's dashboard. It's newer than the others on this list and the documentation shows it, so treat it as promising rather than production-ready until more people have kicked the tyres.

Look at what's actually getting starred this week and a pattern falls out. Almost none of it is about making the model smarter. It's about making everything around the model cheaper, faster or less annoying: compressing what you feed it, keeping its documentation current, giving it hands to use a website with. That tracks with something I say to clients constantly. Most work thrown at frontier models does not need frontier reasoning, and the tooling market has clearly noticed the same thing before most teams did. Nobody's shipping a smarter brain this week. They're shipping plumbing. And the plumbing is where the actual margin lives. A 20% token reduction across a busy agent is real money every single day. A marginally cleverer model is a line item you might not even use correctly.

Is your team optimising the model or the pipes around it?

Tell me in the comments.

Forward this to one person who should be using AI better than they are. Reply with what you built, tried or broke this week. I read every one.

Gareth, founder of The Anthropic Stack (theanthropicstack.com)

P.S. Shipping something this week? Railway gives Anthropic Stack readers $20 in credits to start: theanthropicstack.com/railway. It is a referral link, so I earn a small cut if you stay. I only point you at tools I run myself.