antonvishal/browserbase-router

Browserbase Router is a tool that decides which web browsing capability a task actually needs, so that a program never spends more resources than necessary to complete it. When a task comes in, it is checked against a set of rules, and a model is only consulted when those rules cannot confidently decide on their own. Based on that assessment, the task is routed to one of four options: a plain chat reply that needs no web access at all, a search that discovers relevant links, a fetch that reads one already-known page, or a full browser agent session reserved for tasks that require clicking around, logging in, or multi-step research. The project is careful about keeping this decision safe and predictable. The underlying AI model is only used to judge what kind of capability a task requires, never to directly pick a specific paid Browserbase feature itself. If the model is unsure or only confident enough for a plain chat answer, the system defaults to the cheaper search option rather than guessing at something more expensive. A simple page fetch can escalate to a full agent session, but only if the fetch fails due to a connection or server problem, not for ordinary errors like being rate limited or getting an empty result. The repository is a private monorepo containing a reusable npm package called browserbase-router, plus an example application that connects the router to Slack through an assistant called Eve, complete with live progress updates and links back to the underlying agent sessions. Running it requires Bun version 1.3.11 and Node.js 24. The project explicitly states that it is not a security or authorization system on its own, meaning any application using it still has to decide separately who is allowed to submit tasks and which browser actions are permitted. It is released under the MIT license.