AutoBrowse Agent Browser

Build reliable browser automation skills through iterative experimentation using agent-browser instead of Browserbase browse. Inner agent execution routes through ACPX, not direct Anthropic SDK/API keys. See README.md for full operator guide.

Install from skills.sh

npx skills add https://github.com/t0dorakis/theo-pi --skill autobrowse-agent-browser

Preview without installing:

npx skills add https://github.com/t0dorakis/theo-pi --skill autobrowse-agent-browser --list --full-depth

Source directory

Repo source lives at:

skills/autobrowse-agent-browser

If installed globally, set this to the installed skill directory before running harness scripts:

export AUTOBROWSE_AGENT_BROWSER_DIR="${AUTOBROWSE_AGENT_BROWSER_DIR:-$PWD/skills/autobrowse-agent-browser}"

Setup

export AUTOBROWSE_AGENT_BROWSER_DIR="${AUTOBROWSE_AGENT_BROWSER_DIR:-$PWD/skills/autobrowse-agent-browser}"
cd "$AUTOBROWSE_AGENT_BROWSER_DIR"
npm install
command -v agent-browser || npm install -g agent-browser
agent-browser install # downloads browser runtime; may need extra disk/network access

No ANTHROPICAPIKEY is required. Configure ACPX/Claude Code auth the same way this repo's ACPX delegation does.

Workflow

1. Create workspace:

mkdir -p ./autobrowse/tasks ./autobrowse/traces ./autobrowse/reports

1. Create task:

TASK=<short-kebab-name>
mkdir -p ./autobrowse/tasks/$TASK
cp "$AUTOBROWSE_AGENT_BROWSER_DIR/references/example-task.md" ./autobrowse/tasks/$TASK/task.md
$EDITOR ./autobrowse/tasks/$TASK/task.md

1. Run one evaluation:

node "$AUTOBROWSE_AGENT_BROWSER_DIR/scripts/evaluate.mjs" --task "$TASK" --workspace ./autobrowse
# optional: --agent claude --timeout 1200 --deny-all

1. Read trace:

cat ./autobrowse/traces/$TASK/latest/summary.md 2>/dev/null || ls -t ./autobrowse/traces/$TASK/run-*/summary.md | head -1 | xargs cat

1. Update only ./autobrowse/tasks/$TASK/strategy.md with one concrete improvement, then rerun.

1. Stop after 3-5 iterations or when 2 of last 3 runs pass.

agent-browser command rules for learned strategies

Use these primitives in generated strategies:

agent-browser open <url>
agent-browser wait --load networkidle
agent-browser snapshot -i
agent-browser click @e1
agent-browser fill @e2 "value"
agent-browser keyboard type "text"
agent-browser press Enter
agent-browser scroll down 500
agent-browser screenshot traces/task/screenshots/step-01.png
agent-browser get url
agent-browser get title
agent-browser get text body
agent-browser eval 'document.title'
agent-browser close

Refs are @eN, not Browserbase [0-5]. Always snapshot after navigation and after each DOM-changing action. Harness injects a unique agent-browser --session ... name for parallel-safe runs.

Graduation

When stable, write a self-contained skill to either:

• skills/<task-name>/SKILL.md if source should be committed in this repo.
• ~/.agents/skills/<task-name>/SKILL.md for local-only use.

Graduated skill must include purpose, trigger conditions, exact workflow, site-specific gotchas, failure recovery, and expected JSON output.

Pitfalls

• Do not use Browserbase browse commands or [X-Y] refs.
• Do not edit task.md during iteration except initial creation.
• Do not repeat failed actions; read trace, form one hypothesis, patch strategy.md.
• Harness uses ACPX --approve-all by default so delegated Claude can browse and write trace artifacts; run only in trusted workspaces.
• Treat ./autobrowse/traces/ as sensitive; snapshots and messages may contain page text or authenticated data; repo root autobrowse/ is gitignored as a backstop.
• Prefer deterministic agent-browser eval extraction when visible text is insufficient.