Browser automation · agent discipline

Pilot Pack

v1.0 MIT starter VS Code · Cursor · Claude Code

Giving an AI coding agent control of a browser is easy. Giving it control without it wandering off and clicking the wrong thing is the hard part. Pilot Pack is the permission-first operating model that makes the second one routine.

License
MIT starter
Tiers
$0 / $29 one-time
Requires
Browser MCP installed separately
Adopt in
An afternoon

What it is

The discipline layer, not the engine.

The patterns in Pilot Pack are six months of trial and error driving real browsers with agents — compressed into something you can adopt in an afternoon. It is the operating model, the prompts, and the hard-won patterns — the part nobody hands you.

01

Permission-first by default

Every browser action is a proposal the operator can see and approve before it runs. The agent gets the capability; you keep the brakes. That is the whole difference between useful and dangerous.

02

Portable across editors

The Pro toolkit's three skills drop into VS Code, Cursor, Claude Code, or Continue. They are patterns, not a plugin tied to one editor that breaks on the next update.

03

Patterns that survive real forms

The multi-row table fill. The iframe pierce. The failure modes and how to read them. The things a clean demo never shows you, written down because they cost real sessions to learn.

The Pro toolkit

What the paid tier holds.

The free starter gets an agent driving a browser under supervision. The Pro toolkit is the compressed experience on top.

Skills

Three portable skills

Drop-in agent skills for the browser-automation tool surface — the moves you reach for on every job.

Playbook

The TIPS playbook

The patterns that took real sessions to find — written so the next person does not have to pay for them again.

Reference

Tool-surface reference

Every tool in the browser-automation surface — what it is for, what it returns, and where it bites.

Who it's for

If this sounds like you, it will fit.

  • Anyone running an AI coding agent — Claude Code, Cursor, Copilot, Continue — who wants it to drive a browser: fill forms, pull data, run a workbook, under supervision.
  • Developers who already use Playwright debug Chrome and want their agent to use it well, not flail at it.
  • Anyone who handed an agent a browser once, watched it go sideways, and wants the discipline instead of the chaos.

Plainly

What it does, and what it doesn't.

What it does

  • Gives an agent a permission-first operating model for the browser
  • Walks the setup for browser automation over MCP
  • Ships three portable skills (Pro)
  • Includes the TIPS patterns playbook (Pro)
  • Includes a full tool-surface reference (Pro)

What it doesn't

  • It does not include the Browser-Automation MCP server — you install that (Playwright + CDP) separately
  • It is not a replacement for Browser-Use or Stagehand — it is the discipline layer on top, not the engine
  • It does not promise every pattern works on every form shape on the web
  • It is not an autonomous agent — the operator stays in the loop by design

Get it

Two ways in.

Start free and get an agent driving a browser under supervision today. The Pro toolkit adds the six-months-compressed part.

Free starter

Starter

$0 · MIT

Enough to give an agent supervised browser skill today.

  • The permission-first operating model
  • Setup mechanics for browser automation via MCP
  • Reference prompts to start from
  • Cloned from a public template
Clone the starter on GitHub

No account. Clone it and it is yours.

Hand your agent the browser. Keep the brakes.

Clone the free starter, wire it to your agent, and run a supervised job. The Pro toolkit is there once you want the patterns nobody hands you.