chrome-bridge lets any AI agent drive your real logged-in Chrome
Local-only, MIT, no Playwright re-login. Plus GPT-6 Astra nearly solves ARC-AGI-3 for ~$19K a run, and Codex quietly makes it the default.
Copy markdown
Your session, not a clean one
chrome-bridge, out today under MIT, drives the Chrome you're already logged into — SSO, cookies and all — instead of launching a fresh Playwright browser. A zero-dependency Node CLI talks over WebSocket to an unpacked extension; agents run `node cli.mjs snap` or POST to 127.0.0.1:9333 to click by element ref, fill forms, and read network traffic. Works with Claude Code, Cursor, Codex, Kimi and Qwen.
What it unlocks this weekend
Point an agent at flows Playwright couldn't reach without you handing over credentials: scrape a dashboard behind SSO, drive an internal admin tool, or fill a form inside your live session. It binds to localhost only and tags driven tabs with a purple pill and a tab group so you can watch — and kill — it. It is your real session, though, so scope tightly what the agent is allowed to touch.
It's already a category
chrome-bridge is one of a fast-growing cluster of "give your agent your real browser" tools. sderosiaux's chrome-agent ships the same idea as a single 3 MB Rust binary — no Node, no API key — with stable element uids and bot-detection patches. If browser-driving agents are on your roadmap, there's now a menu to pick from.
Astra aces ARC-AGI-3 — for $18,817 a run
OpenAI's GPT-6 Astra scored ~99.9% on ARC-AGI-3 using a "provider adapter" harness that keeps opaque reasoning state between calls — but a single run cost $18,817. On the standard harness it manages 62.7%. The frontier can nearly solve the hardest public reasoning benchmark; it just can't do it at a per-task price you'd pay yet.
Codex quietly makes Astra the default
Codex CLI 0.153.4 makes GPT-6 Astra the default when no model is configured, and adds Amazon Bedrock routes. Run Codex unconfigured and you're now on a $5/$25-per-1M, text-only model — so image inputs stop working and your token bill shifts without you changing a setting.
Elsewhere: a free Flash model on Vercel
InclusionAI's Ling 3.0 Flash Sante is free on Vercel's AI Gateway through Oct 4 via a complimentary model ID — a no-cost way to wire a fast model into an app before you commit any spend.