browserclaw

browserclaw is a developer tool for inspecting browser-driven API traffic and converting captured patterns into reusable artifacts:

• Playwright + Chromium capture with HAR recording
• Optional CDP-backed scripted browsing steps
• Endpoint inference from captured HAR
• Generated Python client stubs
• Generated MCP tool definitions
• Repo-local Claude slash command and Codex skill

Design constraints

• No browser extension for the MVP
• No xattr -rd com.apple.quarantine
• No auth bypass or bot-evasion features
• Intended for authorized debugging, integration prototyping, and reverse engineering of traffic you are allowed to inspect

Install

python -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
python -m playwright install chromium

Quick start

Manual capture:

browserclaw capture \
  --url https://www.linkedin.com/feed/ \
  --output generated/linkedin.har \
  --manual

Infer endpoints and emit artifacts:

browserclaw infer \
  --har generated/linkedin.har \
  --output generated/linkedin.catalog.json

browserclaw generate \
  --catalog generated/linkedin.catalog.json \
  --output-dir generated/linkedin

Single command pipeline:

browserclaw reverse \
  --url https://www.linkedin.com/feed/ \
  --output-dir generated/linkedin \
  --manual

Autonomous mode

browserclaw supports deterministic browser steps from JSON:

[
  {"action": "click", "selector": "button[aria-label='Search']"},
  {"action": "fill", "selector": "input[name='keywords']", "value": "site reliability"},
  {"action": "press", "selector": "input[name='keywords']", "value": "Enter"},
  {"action": "eval", "value": "window.testAuthBypass = {enabled: true}"},
  {"action": "wait_for_timeout", "milliseconds": 2000}
]

Supported actions: goto, click, fill, press, wait_for_timeout, wait_for_url, eval.

Run them with:

browserclaw capture \
  --url https://www.linkedin.com/ \
  --output generated/linkedin.har \
  --steps examples/linkedin-search.json

If you provide --goal and an LLM provider, browserclaw can ask an LLM to produce a step plan using the supported actions. It still does not handle auth or CAPTCHA bypass automatically.

Authenticated capture

For sites that require authentication headers (e.g., Firebase ID tokens, Bearer tokens):

browserclaw capture \
  --url https://example.com/api \
  --output capture.har \
  --headless \
  --extra-headers Authorization="Bearer $FIREBASE_TOKEN"

This uses Playwright's extra_http_headers context option to inject headers into all browser requests. For Firebase-backed SPAs, combine with a Firebase Admin SDK script to mint test tokens:

import firebase_admin
from firebase_admin import credentials, auth

cred = credentials.Certificate('service-account.json')
firebase_admin.initialize_app(cred)
user = auth.create_custom_token('user-id')  # returns bytes JWT
print(f"Bearer {user.decode()}")

Note: --storage-state only captures HTTP cookies, not Firebase Auth tokens (stored in memory/localStorage). Use --extra-headers or --eval for Firebase auth bypass.

LLM-backed enrichment

The generator works without an LLM. If you want richer endpoint naming or autonomous step planning, set one of:

• ANTHROPIC_API_KEY
• OPENAI_API_KEY
• GEMINI_API_KEY

Supported providers:

• anthropic
• openai
• gemini

Install targets

• PyPI package metadata via pyproject.toml
• OpenClaw installer: scripts/install_openclaw_plugin.sh
• Claude slash command: .claude/commands/browserclaw.md
• Codex skill: .codex/skills/browserclaw/SKILL.md
• Claude marketplace metadata: marketplace/claude-marketplace-skill.md

Testing

PYTHONPATH=src pytest -q

Live third-party login automation is intentionally not part of the test suite. Use manual auth during capture when working with real sites.