A local-first browser runtime for AI agents.

Useful browser capability without ambient machine authority.

AI agents can inspect dynamic applications, fill forms, download files, capture evidence, and complete operational workflows. Conventional automation often inherits more authority than the task requires: ambient browser profiles, persistent cookies, arbitrary origins, unrestricted JavaScript, local files, broad network reach, or an unauthenticated remote control port.

Cockroach Browser separates browser capability from ambient machine authority. A host creates an explicit session with a purpose, admitted origins, allowed actions, allowed effects, and finite budgets. The runtime then provides semantic page references, browser interactions, screenshots, PDFs, traces, network observations, audits, and hash-linked receipts inside that session.

One package, several explicit control surfaces.

Version 0.3.0 ships an embedded TypeScript SDK, authenticated loopback daemon, typed client, command-line interface, observation-first MCP server, Docker deployment, local dashboard, explicit browser providers, runtime-owned persistent profiles, authenticated jobs and workers, team-scoped roles, and adapters for Maqam, Qarinah, Cockroach Crawler, and ProductLoop OS.

The annotated v0.3.0 source-derived capability registry contains 94 entries: 88 directly available surfaces and 6 host-backed adapters. This is an implementation inventory, not a performance or security score.

Implementation-backed and open for technical review.

This is an implementation-backed technical white paper for Cockroach Browser 0.3.0. The paper has not undergone independent peer review or independent security certification.

Published v1.1: doi:10.5281/zenodo.21850760. Persistent paper series: doi:10.5281/zenodo.21701791. Version 1.0 remains preserved at doi:10.5281/zenodo.21701792.

The software is licensed under AGPL-3.0-or-later. The paper is licensed under Creative Commons Attribution 4.0 International.