Your AI agent shouldn't need to relearn the web every run

What it does Webcmd is a self-learning browser infrastructure for AI agents. It watches an agent fumble through a site–clicking, typing, extracting–then memorizes the navigational context and compiles it into reusable, deterministic commands. The idea is to stop paying an LLM to rediscover the same login flow on every single run.

The interesting bit Instead of treating every task like a fresh Playwright script, Webcmd builds a layered memory: raw exploration at the bottom, a learned sitemap in the middle, and a compiled CLI adapter on top. Each layer collapses variance for the next, turning repeated browser chaos into something closer to a typed API.

Key highlights

• Ships as a Node.js CLI with a plugin-based adapter system; the core includes no site adapters, so you bring your own plugins.
• Supports authenticated sessions via cookie-jar Profiles and parallel Sessions for multi-agent work.
• Aims to cut browser-agent token spend by up to 90% by replacing repeated exploration with deterministic, compiled commands.
• Can target authenticated web apps, APIs, and desktop surfaces beyond standard public sites.
• Community plugins already cover research, social media, shopping, and travel sites.

Caveats

• The core package ships bare: most workflows require hunting down and installing a plugin before anything useful happens.
• The hosted Webcmd Cloud option is explicitly unstable and still in active development.

Verdict Worth a look if you’re burning through API tokens on repetitive browser automation and want to turn exploration into reusable infrastructure. Skip it if you need a fully hosted, batteries-included solution today.