BrowserSkill 0.3.0 lets AI agents use your logged-in browser
Tencent's CLI borrows a tab from your real Chrome, now with canvas — while Vercel turns its Sandbox into an eval harness and Hister indexes your files.
Copy markdown
The login wall, gone
BrowserSkill's `bsk` command borrows a tab from your own logged-in Chrome or Edge, so an agent inherits your cookies and 2FA instead of stalling at a sign-in screen. It works with Claude Code, Cursor, Codex and any shell-capable agent, is MIT-licensed, and cleared 1,300 GitHub stars within a day.
Canvas is finally clickable
0.3.0 adds canvas element detection with typed visual references and region screenshots, plus full-page capture, scroll-to-element and real mouse-wheel input. That lets an agent see and click Figma boards, maps and other canvas-drawn UIs the DOM never exposed.
Vercel Sandbox becomes an eval harness
Harbor 0.22.0 now runs Terminal-Bench agent evals inside isolated Vercel Sandbox microVMs, so you can score an agent without standing up your own infra. The same changelog adds sub-second CLI artifact deploys and pulling agent skills straight from a Notion database.
Elsewhere: a private search index you own
Hister is an AGPLv3 single binary that indexes the pages you visit and the files you keep, searchable from a web UI, the terminal, or your agent over MCP. No telemetry and no external calls by default — a weekend self-host that hands your agent a memory of everything you've read.
Heads-up: patch your Datasette
Datasette 1.0a39 and 0.65.5 ship security fixes for a table-permission bypass that could expose rows past your access rules. If you run a public Datasette instance, upgrade now.