What it does

Webcmd is a command-line tool for AI agents that need to work in a real browser. On an unfamiliar site it offers live browser control to inspect, click, type, extract data and capture network calls. On a familiar site it uses a memory layer that records an agent-facing sitemap of pages, states, actions, workflows, APIs, pitfalls and fallback paths. The aim is that agents no longer relearn the same sites on every run. The project says this can cut browser-agent token spend by up to 90%, and it publishes a benchmark comparison on a 100-task browser automation set. Learning is described as quiet and selective, and a memory failure never blocks the task. First access may use a Webcmd Cloud seed, and later learning stays local. Setup needs Node.js 20.6 or later. After installing the package with npm, a skills command adds one skill, webcmd-browser, to Claude, Codex, another supported harness or a custom skills path. Named profiles act as cookie jars so an agent can work in a logged-in session, and separate sessions let parallel agents use their own browser windows. Webcmd Cloud can run commands and browser sessions on hosted infrastructure. The page says it is in active development and not yet stable.

Capabilities

Live browser control to inspect, click, type, extract and capture network calls
Sitemap memory of pages, actions, workflows and pitfalls
Named profiles for logged-in work and separate sessions for parallel agents
Installs one skill, webcmd-browser, for Claude, Codex and other harnesses
Runs sandboxed Playwright-style programs in a browser session
Hosted Webcmd Cloud option, still in development

When to use it

Your agent keeps re-exploring the same websites and spending tokens on it
A task needs a logged-in browser session, such as reading messages or bookmarks
You want several browser agents to run in parallel with isolated sessions