iframer doesn't just browse a website, it learns it: caching endpoints and creating "anchors" for its key elements, so the next visit is instant.
It also covers the basics: secure sessions, headed/headless modes, self-hosted captcha solving.
Built for real engineering work, not one-off research, and the benchmarks below prove it.
Hard tasks: many pages, real finding-your-way
Everest → Gojo, 6 hops, find each link
login → sort → cart → checkout
GitHub: look + pull 4 facts
Shorter bar wins. Where the others dump a whole accessibility tree into the model’s context, iframer returns a capped, filtered snapshot, so on the GitHub page it spends 986 tokens where Playwright spends 19,906. See the full benchmark →
Runs locally. No Docker required. Works with Claude Code, Codex, any MCP agent.
What it does
Browser modes
iframer auto-selects and auto-escalates: headless → binary-headful → docker-headful. See architecture for how that decision actually works.
How it works
What's inside
Full breakdown of the daemon, the flow, and the stack: architecture. Everything an agent actually calls: agent docs.
Open source · MIT license · 2026