Tsaagan

Browser automation that stops after a click is useless for agents that need to know the form submitted successfully before filing the next request. Tsaagan wraps that verification step into its core loop: the agent perceives the page state, takes an action, and confirms the outcome before the pipeline advances. The tool exposes this capability through an MCP server for agent frameworks that speak that protocol, plus dedicated SDKs for JavaScript and Python, so the integration surface matches wherever your agent runtime lives.

The verify-first model is the differentiating design choice. Standard Playwright scripts — and most wrapper libraries built on top of it — assume success unless an exception fires. Tsaagan inverts that: confirmation is mandatory, not optional. For agents driving multi-step authenticated workflows — login, navigate, fill, submit, check — this means the agent doesn’t proceed on an assumption. That distinction matters most when the downstream consequence of a silent failure is a corrupted pipeline or a missed transaction.

The repo is MIT licensed and self-hosted only; there is no managed cloud offering. That means you own the infrastructure, the scaling, and the debugging. At small agent counts, a single self-hosted instance handles the load. As concurrent agent sessions multiply, teams will hit the limits of early-stage software before they hit the limits of Playwright itself — at which point migration to a more mature automation backend becomes the practical path. The project ships an acceptable use policy and a roadmap file, which suggests the maintainer is thinking about governance, but community size and issue velocity are indicators of where the project sits in its lifecycle.