Webcmd

Webcmd is browser infrastructure for AI agents that learns websites as agents use them and turns that knowledge into reusable commands. The problem it targets is that browser agents repeatedly rediscover the same pages with screenshots, DOM snapshots and fresh reasoning, which is slow, costly and fragile. Webcmd keeps what an agent observes, including pages, actions, APIs, pitfalls and fallback paths, and promotes it into a durable command with named inputs and structured output. The workflow has four layers. Control lets an agent inspect, click, type, extract and capture network calls in a real browser. Remember records a sitemap of observed pages, states and workflows. Author builds a reusable site adapter with explicit arguments and strategy tags. Execute runs the resulting commands from the command line, for example a command that returns the top Hacker News stories as JSON. Webcmd chooses the cheapest reliable path for each task, whether a public endpoint, authenticated cookies, intercepted requests, the page interface or a local tool. It installs through npm and adds skills to a user's coding agent, and the source is on GitHub. The target audience is developers who run browser agents repeatedly against the same sites and want lower token use and more predictable results. The benchmark figures on the home page are labeled as illustrative placeholders. The site publishes no pricing.