agentbrowse: The Terminal Interface Your AI Coding Agent Has Been Waiting For

Let's be honest: AI coding agents are incredibly powerful when they're working with command lines and code files, but ask them to click through a web interface and things get... messy.

That's the gap agentbrowse fills. It's a CLI tool that wraps Playwright into a simple, agent-friendly interface — one that speaks the language of AI coding assistants like Claude Code, Codex, Cursor, and others.

The Problem With Web UIs and AI Agents

Traditional browser automation assumes a human is driving. Humans can handle CAPTCHAs, parse visual layouts intuitively, and click exactly where they need to on a pixel-perfect button. AI agents? Not so much. They're great at executing precise commands but struggle with the ambiguity of visual interfaces.

agentbrowse flips this relationship. Instead of an agent wrestling with a browser, it gets a clean API: open a URL, read the page as markdown, click by text or number, fill forms, and more — all through terminal commands that return structured data.

Getting Started in Minutes

Installation is straightforward:

npm install -g agentbrowse
npx playwright install chromium

That's it. The first run automatically downloads the Chromium browser, and you're ready to go.

The core workflow is beautifully simple:

• agentbrowse open <url> — navigate to a page
• agentbrowse read — get the current page as clean, token-bounded markdown
• agentbrowse links — see numbered, followable links
• agentbrowse click "Sign up" — click by visible text or link number
• agentbrowse read --json — structured output for programmatic parsing

Persistence: The Secret Sauce

What makes agentbrowse genuinely useful is its persistent session model. A background browser daemon runs continuously, keeping your page state alive between commands. This means you can open a site, authenticate, browse to a specific page, and then hand off to your agent — all without rebuilding state each time.

Sessions are isolated by ID, with their own cookies and authentication state. The login command even opens a real browser window for you to handle MFA and CAPTCHAs, then saves that authenticated session for headless reuse.

Making Your Agent Reach for It Automatically

Here's where things get clever. You can install agentbrowse as a "skill" that your coding agent recognizes automatically:

npx agentbrowse skill --global

This writes the appropriate configuration file for whatever agent you're using — SKILL.md for Claude Code, .cursor/rules for Cursor, AGENTS.md for Codex — so your AI assistant knows to reach for browser automation when web interactions come up.

No more elaborate prompts explaining how to use a browser. The capability is just... there.

Structured Output for Machine Consumption

Every command accepts --json for machine-readable output, and errors follow a consistent format with specific exit codes (2 for usage errors, 3 for navigation failures, 4 for target-not-found, 5 for daemon issues). This makes agentbrowse scriptable and reliable for CI/CD pipelines and automated workflows.

Site Manifests: Domain-Specific Superpowers

For complex or frequently-used sites, you can define a site manifest file that exposes named, high-level commands:

agentbrowse --site ./github.agent.json create-issue "Bug in login flow"

Manifests declare pages, selectors, and command sequences with variable interpolation — essentially creating custom "skills" for specific websites.

The Bigger Picture

agentbrowse represents a shift in how we think about AI agent capabilities. Rather than hoping AI vision models can approximate human-browser interaction, we give agents tools designed for their strengths: structured commands, parseable output, and persistent state.

For developers building agent-powered workflows, this is a practical tool that fills a real gap. For those exploring AI-assisted development, it's a glimpse at how agent-tool interactions might evolve — away from visual interfaces designed for humans, toward programmatic interfaces designed for machines.

Whether you're automating web testing, building AI workflows, or just curious about the future of agent-tool interaction, agentbrowse is worth exploring. It's MIT-licensed, actively maintained, and runs entirely locally — no cloud dependencies or external services required.

The terminal is back, and now even your AI can use it.