The AI browser agent that returns a verdict

The honest position from TestMu AI: the LLM is not deterministic, but the validation contract is. The model decides how to reach an element, and that path can vary run to run, but it does not get to decide the test passed. A pass is granted only when the expected state is verified through explicit evidence: DOM state, stable selectors, accessibility labels, URL changes, network responses, screenshots, console logs, or your own assertions. In CI you key the pipeline off the evidence-backed exit code: 0 when the expected state is verified, 1 on a failed assertion, 2 on an error such as auth, and 3 on timeout. That is what makes a non-deterministic model safe to gate on.

Yes. Because the verdict is machine-readable, AI coding agents can close their own loop. Run Kane CLI with the --agent flag and it streams NDJSON, one typed event per line, ending in a run_end event carrying the verdict, an evidence summary, extracted values, and a dashboard link. Claude Code, Cursor, Codex, and Gemini CLI parse that stream to learn whether the UI they generated actually works, then fix the regression before you open the browser. Point your agent at the guide at testmuai.com/kane-cli/agents.md and it installs, authenticates, and verifies in a real browser on its own.

When one Chrome on your laptop is not enough, the same objective runs on Browser Cloud, the browser infrastructure built for AI agents. It provisions real, full-featured Chrome sessions on demand and runs many in parallel, with a built-in tunnel to reach localhost and staging, automatic session video and logs for debugging, and persistent login state across runs, all on the same cloud that powers 1.5 billion tests a year for 18,000+ enterprises. Start in your terminal, then scale out without a rewrite. The Browser Cloud documentation and the browser infrastructure for AI agents post cover the setup.

Yes. Beyond the --agent stream, TestMu AI exposes browser automation over the Model Context Protocol, so any MCP-compatible agent can navigate, interact, and read the page as structured tool calls inside its own task loop, with no codegen step. See the browser automation MCP server for the tool surface. Use the MCP server when the agent needs to act on the web in real time, and the CLI when you want a verified pass or fail to gate on.

Security comes from two design choices. First, the agent acts only on a real Chrome browser over the DevTools Protocol, limited to actions a real user could take, with no injected JavaScript that fakes UI state, so every run leaves an honest evidence trail you can audit. Second, when you scale on Browser Cloud the infrastructure carries SOC 2 Type II, ISO 27001, HIPAA, and GDPR compliance, with credentials supplied through environment variables and CI secrets rather than hardcoded. Treat stealth and CAPTCHA handling as best-effort, never a guarantee, and keep your access key out of source control.

Traditional browser automation anchors every action to a brittle selector or XPath, so a renamed label or a new CSS class breaks the run and a human fixes the script. A browser agent anchors to intent, the user-facing element like "the Submit button", and re-resolves it when the page shifts, autohealing cosmetic drift and pushing through unexpected modals up to 50 steps per flow. You write what you want, not the path to get there. Any completed run still exports to native Playwright, and you can scale framework suites on the Automation Cloud grid, so there is no lock-in.

Yes, Kane CLI is free to install and free to run against your own Chrome, so you can prove out the LLM-perception-plus-evidence workflow at zero cost. The Starter tier adds 100 credits with no credit card. Cloud runs on the TestMu AI grid are billed against your plan only when you scale to remote browsers, geo coverage, or parallel sessions. Start free, verify a real journey with shareable evidence in under five minutes, and pay only when you need scale.