The agent-browser alternative that returns a verdict

agent-browser is an open-source browser automation CLI for AI agents from Vercel Labs, published under Apache-2.0 and shipped as a native Rust binary with no Node.js required at runtime. Its core loop is open, snapshot, act on accessibility-tree refs such as @e1, then re-snapshot. It is genuinely good software: 39,293 GitHub stars and roughly 1.13 million npm downloads a week as observed on July 27, 2026, and its snapshots are deliberately token-efficient.

Three differences matter. First, the input: you write an objective in natural language instead of sequencing imperative commands, so there are no refs to track and no selectors to maintain. Second, the contract: agent-browser documents no assert or expect command, so your agent decides what counts as working by reading page state, while Kane CLI grants a pass only when evidence confirms the expected state. Third, the output: Kane CLI emits an NDJSON event stream ending in a run_end verdict plus exit codes 0, 1, 2 and 3, so CI can gate on it directly.

In several places, and it is worth being straight about them. It is Apache-2.0 licensed with the full source public, so it is free at any volume with no account and no license key. It ships as a single native Rust binary that needs no Node.js runtime, and it publishes builds for macOS ARM64 and x64, Linux ARM64 and x64, and Windows x64. It documents skills for a wider set of clients than we do, including Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Goose, OpenCode and Windsurf, plus a built-in MCP stdio server. It also bundles axe-core accessibility audits, Web Vitals and React DevTools introspection. If you want a free, local, source-available browser driver and you are happy owning the verification logic yourself, it is a strong choice.

Yes. Run any objective with the --agent flag and Kane CLI suppresses the interactive UI and streams typed NDJSON, one JSON object per line, ending in a run_end event carrying status, summary, extracted values, token usage and a dashboard link. Claude Code, Codex CLI and Gemini CLI can install Kane CLI as a skill. Point your agent at testmuai.com/kane-cli/agents.md and it will install, authenticate and verify in a real browser on its own.

Yes, and this is the clearest reason teams add it alongside a browser driver. Combine --agent with --headless so no display server is needed, then key your pipeline off the exit code: 0 when the expected state is verified, 1 on a failed assertion, 2 on an error such as an auth failure, and 3 on timeout. Kane CLI documents native integration for GitHub Actions, GitLab CI, Jenkins and Bitbucket Pipelines. agent-browser documents an exit-code contract for its doctor command, but not a general table for its automation commands, so a pipeline built on it needs you to define the pass condition yourself.

agent-browser refs are scoped to a snapshot. Its documentation states that refs are invalidated when the page changes and instructs the agent to take a fresh snapshot before retrying, which is correct and predictable, but it means the recovery loop is your agent's job and costs another round trip. Kane CLI anchors to the user-facing element rather than a ref, so autoheal re-resolves intent when a label is reworded or a class changes, and a run pushes through unexpected modals and redirects for up to 50 steps while the pass condition stays exactly the same.

agent-browser delegates scale to third-party providers you hold accounts with, including Browserless, Browserbase, Browser Use, Kernel and AWS Bedrock AgentCore. Kane CLI scales onto first-party infrastructure instead: the same objective runs on Browser Cloud, which provisions real Chrome on demand with a built-in tunnel to reach localhost and staging, session video and logs, and persistent login state. For mobile, TestMu AI runs a real device cloud with 10,000+ real devices. Read the browser infrastructure for AI agents post for the architecture.

Yes. The Free plan costs $0 and ships 200 credits, and it runs against your local Chrome, so you can prove out the whole workflow without a credit card. Paid plans add scale rather than unlocking the basics: Starter at $19 a month with 2,000 credits, Pro at $99 a month with 10,000 credits, and Enterprise on request. agent-browser is free at any volume because it is Apache-2.0, though its optional chat mode requires an AI_GATEWAY_API_KEY and bills model inference to you, and its cloud providers bill separately. Plans are on the pricing page.