The pattern shaping this week's trending repos is agents getting better at operating in the world rather than just reasoning about it. Chrome's official DevTools team shipped an MCP server so agents can debug live browsers. Alibaba released a browser automation library that skips screenshots entirely. And a skill that trims agent verbosity by 65% crossed 84,000 stars — a signal that token efficiency has become a genuine engineering priority, not just a cost footnote.
Chrome DevTools Gets an MCP Server
The Chrome DevTools team published an official MCP server that lets coding agents control and inspect live Chrome instances. Beyond basic navigation, agents can now record performance traces, analyze network traffic, track console output with source-mapped stack traces, run Lighthouse audits, and take heap snapshots. The integration works across Claude, Cursor, VS Code Copilot, and Gemini. What makes this different from prior browser-automation MCP tools is provenance: this comes from the DevTools team itself, which means it's built on the same reliable Puppeteer foundation and is likely to stay current as the DevTools protocol evolves. Teams building agents that need to debug real web apps in CI now have a well-maintained, cross-platform path to do it.
github.com/ChromeDevTools/chrome-devtools-mcp
Alibaba's Page Agent: Browser Automation Without Screenshots
Alibaba open-sourced Page Agent, a client-side JavaScript library that lets agents drive web interfaces through natural language by reading the DOM directly rather than interpreting screenshots. That distinction matters: screenshot-based approaches require multimodal models and tend to break on dynamic UIs; Page Agent operates on document structure, which is more stable and works with any capable text model. Integration requires one line of JavaScript. An optional Chrome extension handles multi-page task chains, and an MCP server in beta allows external agent control. The project is pitched at SaaS platforms adding AI copilots, CRM and ERP systems needing form automation, and accessibility use cases where voice commands need to navigate complex interfaces.
Caveman: 65% Fewer Output Tokens, No Capability Trade-off
Caveman is a prompt-based skill installable across 30-plus agent platforms — Claude Code, Cursor, Cline, Copilot, and others — that cuts agent output tokens by around 65% by eliminating filler language and enforcing terse communication. It never modifies code, commands, or error messages, only prose. Six compression levels from "lite" to "ultra" are switchable mid-session, and a companion command rewrites memory files to reduce input tokens across future sessions. The project cites research suggesting brevity constraints can improve model accuracy alongside reducing cost, and benchmarks show 22 to 87 percent savings depending on task type. With 84,000 stars and zero backend dependencies, it's the kind of purely local, no-trust-required tool that spreads fast through developer communities.
github.com/JuliusBrussee/caveman
Agent Skills Becomes an Open Standard
The agentskills/agentskills repository — originally developed at Anthropic and now maintained as an open specification — defines a lightweight format for packaging specialized knowledge into portable, reusable agent plugins. A skill is a folder with a SKILL.md file; agents load only names and descriptions initially, then pull full instructions when a skill is relevant to the current task. The progressive-disclosure model keeps context overhead low while still allowing agents to carry a large library of expertise. The spec has been adopted across multiple agent products and now has 22,400 stars, suggesting it's gaining traction as the de facto way to distribute agent capabilities across platforms rather than building platform-specific integrations each time.