Why Your AI Coding Assistant Can't Use the Web (And What Changes Everything)

Here's a truth nobody talks about openly: AI coding agents are absolute wizards when they're working with files and terminal commands, but hand them a web browser and watch the magic evaporate.

They're great at writing code, following precise instructions, and executing commands. But ask them to navigate a login form, click through a settings page, or fill out a web-based interface? Things fall apart pretty quickly.

That's exactly why agentbrowse exists. It's a command-line tool that wraps Playwright into something AI agents can actually work with — a clean, predictable interface that speaks their language.

The Core Problem

Traditional browser automation tools were built for humans. Humans can read visual layouts, click on buttons that look right, and handle those annoying CAPTCHAs that pop up. We're adaptable. We improvise.

AI agents? They're not so flexible. They execute precise instructions beautifully but stumble when things get ambiguous — and web interfaces are nothing if not ambiguous.

Agentbrowse flips this entirely. Instead of forcing an AI to wrestle with a graphical interface, you give it a straightforward API. Open URLs, read page content as markdown, click by text or link number, fill forms — all through simple terminal commands that return clean, structured data.

Getting Up and Running

Installation takes about thirty seconds:

npm install -g agentbrowse
npx playwright install chromium

That's the whole setup. First time you run it, Chromium downloads automatically. Then you're ready to go.

The command set is refreshingly minimal:

• agentbrowse open <url> — go somewhere
• agentbrowse read — grab the current page as markdown
• agentbrowse links — see numbered links you can follow
• agentbrowse click "Sign up" — click by text or link number
• agentbrowse read --json — structured data instead of text

Clean, right?

Sessions That Actually Persist

Here's where agentbrowse gets genuinely clever. Most browser automation tools treat each action like it starts from scratch. Open a page, close the tool, start again — you're back at square one.

Agentbrowse runs a persistent background daemon. The browser stays alive between commands. You can log into a site, navigate to a specific page, then hand everything off to your AI agent without rebuilding state each time.

Each session gets its own ID, with isolated cookies and authentication. The login command is particularly nice — it opens a real browser window so you can handle MFA codes and CAPTCHAs manually, then saves that authenticated session for the daemon to reuse later. No more fighting with two-factor authentication programmatically.

Making Your Agent Use It Automatically

This is the part I find most interesting. You can install agentbrowse as a "skill" that your preferred coding agent recognizes natively:

npx agentbrowse skill --global

This writes the right configuration files for whatever agent you're running — SKILL.md for Claude Code, .cursor/rules for Cursor, AGENTS.md for Codex. Your AI assistant just knows how to use browser automation when web interactions come up.

No more elaborate system prompts explaining how to click through a web interface. The capability is built in.

Built for Scripting

Every command accepts --json for machine-readable output, and errors follow a consistent pattern with specific exit codes. Usage problems return 2, navigation failures return 3, missing targets return 4, daemon issues return 5.

This predictability makes agentbrowse reliable for CI/CD pipelines and automated workflows. You know exactly what to expect when something goes wrong.

Domain-Specific Manifests

For sites you use frequently, you can define a manifest file that exposes named, high-level commands:

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

Manifests let you declare pages, selectors, and command sequences with variable substitution. Think of it as creating custom shortcuts for specific websites — a way to encapsulate complex workflows into a single command.

The Bigger Picture

What agentbrowse represents matters more than the tool itself. It's a sign that we're moving away from hoping AI vision models can mimic human browser usage, toward building tools that work with AI's actual strengths.

Structured commands. Parseable output. Persistent state between operations.

These are interfaces designed for machines, not humans. And that's exactly what the next generation of AI tooling needs.

Whether you're automating web tests, building agent-powered workflows, or just exploring where this technology is heading, agentbrowse is worth some time. It's MIT licensed, actively developed, and runs entirely on your machine — no external services, no cloud dependencies.

The terminal never really went away. Now even your AI can use it.