Each command prints JSON to stdout and uses semantic exit codes (0 success, 1 error, 2 can't connect, 3 timeout), so they compose naturally with pipes and &&.

I built it because I wanted AI agents to be able to do everything Chrome DevTools can do, not just the basics. Existing options were either too narrow (MCP browser tools cover click/navigate/fill but not network interception or heap snapshots) or too heavy (Puppeteer/Playwright are great but they're libraries — an agent has to write and run a script rather than just call a command).

There's no session to manage and no browser object to keep alive. An agent just runs hubcap goto --wait https://example.com && hubcap text '#price'.

It also includes a Claude Code skill. The skill only loads command categories into context and pulls full docs on demand via hubcap help, so an agent gets coverage of 118 commands without the context cost.

Source: https://github.com/tomyan/hubcap Docs: https://hubcap.tomyandell.dev