Browser Automation

Browser automation for AI agents using PinchTab — a high-performance Chrome bridge with HTTP API.

What is PinchTab?

• Standalone HTTP server — Control Chrome via HTTP API
• Token-efficient — 800 tokens/page with text extraction (vs 10,000+ for screenshots)
• Multi-instance — Run multiple parallel Chrome processes with isolated profiles
• Headless or Headed — Run without window or with visible Chrome
• Self-contained — 12MB binary, no external dependencies
• MCP integration — Native SMCP plugin for Claude Code

Quick Start

Installation

# macOS / Linux
curl -fsSL https://pinchtab.com/install.sh | bash

# npm
npm install -g pinchtab

# Docker
docker run -d -p 9867:9867 pinchtab/pinchtab

Start Server

# Terminal 1: Start PinchTab server
pinchtab
# Server runs on http://localhost:9867

Basic Commands

# Navigate
pinchtab nav https://pinchtab.com

# Wait 3 seconds for accessibility tree
sleep 3

# Get interactive elements
pinchtab snap -i -c

# Extract text (token-efficient)
pinchtab text

# Click element by ref
pinchtab click e5

# Fill input
pinchtab fill e3 "user@example.com"

---

Core Concepts

Instance

A running Chrome process. Each instance has isolated state.

# Create headless instance
pinchtab instances create --mode=headless

# Create headed instance (visible window)
pinchtab instances create --mode=headed

# List instances
pinchtab instances list

# Stop instance
pinchtab instances stop <instance-id>

Profile

Browser state (cookies, history, localStorage). Log in once, stay logged in.

# Create instance with profile
pinchtab instances create --profile=work

# Profile persists across restarts

Tab

A single webpage. Each instance can have multiple tabs.

# Open new tab
pinchtab tabs open https://example.com

# List tabs
pinchtab tabs list

# Close tab
pinchtab tabs close <tab-id>

---

Token-Efficient Patterns

The 3-Second Wait Rule

Critical: Chrome's accessibility tree takes ~3 seconds to populate after navigation.

# ❌ Too fast - empty tree
pinchtab nav https://example.com
pinchtab snap
# Returns: {"count": 1, "nodes": [{"ref": "e0"}]}

# ✅ Wait 3 seconds
pinchtab nav https://example.com
sleep 3
pinchtab snap
# Returns: {"count": 2645, "nodes": [...]}

Optimal Extraction Pattern

# Navigate + wait + filter (14x token savings)
curl -X POST http://localhost:9867/navigate \
  -d '{"url": "https://example.com"}' && \
sleep 3 && \
curl http://localhost:9867/snapshot | \
jq '.nodes[] | select(.name | length > 15) | .name' | \
head -30

Why this works:

1. Navigate + wait ensures full accessibility tree
2. jq filter extracts text nodes only
3. length > 15 filters buttons/labels
4. head -30 limits output

Token comparison:

• Exploratory approach: ~3,800 tokens
• Pattern-driven: ~270 tokens
• Savings: 14x

---

HTTP API Reference

Base URL

http://localhost:9867

Instances

# Create instance
TAB=$(curl -s -X POST http://localhost:9867/instances \
  -d '{"profile":"work","mode":"headless"}' | jq -r '.id')

# List instances
curl http://localhost:9867/instances

# Stop instance
curl -X POST "http://localhost:9867/instances/$TAB/stop"

Navigation

# Navigate to URL
curl -X POST "http://localhost:9867/instances/$TAB/tabs/open" \
  -d '{"url":"https://example.com"}'

# Wait for load
sleep 3

Snapshot

# Full snapshot
curl "http://localhost:9867/instances/$TAB/snapshot"

# Interactive elements only
curl "http://localhost:9867/instances/$TAB/snapshot?filter=interactive"

# With coordinates
curl "http://localhost:9867/instances/$TAB/snapshot?includeCoords=true"

Actions

# Click element
curl -X POST "http://localhost:9867/instances/$TAB/action" \
  -d '{"kind":"click","ref":"e5"}'

# Type text
curl -X POST "http://localhost:9867/instances/$TAB/action" \
  -d '{"kind":"type","ref":"e12","text":"hello"}'

# Press key
curl -X POST "http://localhost:9867/instances/$TAB/action" \
  -d '{"kind":"key","ref":"e12","key":"Enter"}'

# Scroll
curl -X POST "http://localhost:9867/instances/$TAB/action" \
  -d '{"kind":"scroll","direction":"down"}'

Extraction

# Extract text (token-efficient)
curl "http://localhost:9867/instances/$TAB/text"

# Take screenshot
curl "http://localhost:9867/instances/$TAB/screenshot" \
  --output screenshot.png

# Generate PDF
curl "http://localhost:9867/instances/$TAB/pdf" \
  --output page.pdf

# Evaluate JavaScript
curl -X POST "http://localhost:9867/instances/$TAB/evaluate" \
  -d '{"script": "document.title"}'

---

Common Patterns

Pattern 1: Web Scraping

#!/bin/bash
# scrape-headlines.sh

URL=$1
INST=$(curl -s -X POST http://localhost:9867/instances \
  -d '{"mode":"headless"}' | jq -r '.id')

# Navigate and wait
curl -s -X POST "http://localhost:9867/instances/$INST/tabs/open" \
  -d "{\"url\":\"$URL\"}"
sleep 3

# Extract headlines (filter by length)
curl -s "http://localhost:9867/instances/$INST/snapshot" | \
  jq '.nodes[] | select(.name | length > 20) | .name' | \
  head -20

# Cleanup
curl -s -X POST "http://localhost:9867/instances/$INST/stop"

Pattern 2: Form Interaction

#!/bin/bash
# fill-form.sh

INST=$(curl -s -X POST http://localhost:9867/instances \
  -d '{"mode":"headless"}' | jq -r '.id')

# Navigate to form
curl -s -X POST "http://localhost:9867/instances/$INST/tabs/open" \
  -d '{"url":"https://example.com/login"}'
sleep 3

# Get snapshot to find element refs
SNAPSHOT=$(curl -s "http://localhost:9867/instances/$INST/snapshot?filter=interactive")

# Extract refs (example: e5=email, e7=password, e9=submit)
EMAIL_REF=$(echo $SNAPSHOT | jq -r '.nodes[] | select(.name | contains("email")) | .ref')
PASS_REF=$(echo $SNAPSHOT | jq -r '.nodes[] | select(.name | contains("password")) | .ref')
SUBMIT_REF=$(echo $SNAPSHOT | jq -r '.nodes[] | select(.role == "button") | .ref')

# Fill form
curl -s -X POST "http://localhost:9867/instances/$INST/action" \
  -d "{\"kind\":\"type\",\"ref\":\"$EMAIL_REF\",\"text\":\"user@example.com\"}"
curl -s -X POST "http://localhost:9867/instances/$INST/action" \
  -d "{\"kind\":\"type\",\"ref\":\"$PASS_REF\",\"text\":\"password123\"}"

# Submit
curl -s -X POST "http://localhost:9867/instances/$INST/action" \
  -d "{\"kind\":\"click\",\"ref\":\"$SUBMIT_REF\"}"

# Wait for navigation
sleep 3

# Verify login
curl -s "http://localhost:9867/instances/$INST/text" | jq -r '.title'

# Cleanup
curl -s -X POST "http://localhost:9867/instances/$INST/stop"

Pattern 3: Multi-Instance Parallel Processing

#!/bin/bash
# parallel-scrape.sh

URLS=("https://site1.com" "https://site2.com" "https://site3.com")
INSTANCES=()

# Create instances
for i in {0..2}; do
  INST=$(curl -s -X POST http://localhost:9867/instances \
    -d '{"mode":"headless"}' | jq -r '.id')
  INSTANCES[$i]=$INST
done

# Launch parallel jobs
for i in {0..2}; do
  (
    curl -s -X POST "http://localhost:9867/instances/${INSTANCES[$i]}/tabs/open" \
      -d "{\"url\":\"${URLS[$i]}\"}"
    sleep 3
    TITLE=$(curl -s "http://localhost:9867/instances/${INSTANCES[$i]}/text" | jq -r '.title')
    echo "Result $i: $TITLE"
    curl -s -X POST "http://localhost:9867/instances/${INSTANCES[$i]}/stop"
  ) &
done

wait
echo "All complete"

Pattern 4: Visual Regression Testing

#!/bin/bash
# visual-regression.sh

URLS=("https://staging.example.com" "https://production.example.com")
INST=$(curl -s -X POST http://localhost:9867/instances \
  -d '{"mode":"headless"}' | jq -r '.id')

for URL in "${URLS[@]}"; do
  curl -s -X POST "http://localhost:9867/instances/$INST/tabs/open" \
    -d "{\"url\":\"$URL\"}"
  sleep 3

  # Take screenshot
  FILENAME=$(echo $URL | sed 's/[^a-zA-Z0-9]/_/g').png
  curl -s "http://localhost:9867/instances/$INST/screenshot" \
    --output "$FILENAME"
  echo "Saved: $FILENAME"
done

curl -s -X POST "http://localhost:9867/instances/$INST/stop"

Pattern 5: Session Persistence

#!/bin/bash
# persistent-session.sh

# Create instance with named profile
INST=$(curl -s -X POST http://localhost:9867/instances \
  -d '{"profile":"myaccount","mode":"headless"}' | jq -r '.id')

# Login once
curl -s -X POST "http://localhost:9867/instances/$INST/tabs/open" \
  -d '{"url":"https://example.com/login"}'
sleep 3
# ... perform login ...

# Stop (cookies saved to profile)
curl -s -X POST "http://localhost:9867/instances/$INST/stop"

# Later: Resume with same profile
INST2=$(curl -s -X POST http://localhost:9867/instances \
  -d '{"profile":"myaccount","mode":"headless"}' | jq -r '.id')

# Already logged in!
curl -s -X POST "http://localhost:9867/instances/$INST2/tabs/open" \
  -d '{"url":"https://example.com/dashboard"}'

---

MCP Integration

PinchTab provides an SMCP plugin for native Claude Code integration.

Setup

# Set plugin directory
export MCP_PLUGINS_DIR=/path/to/pinchtab/plugins

# Restart Claude Code to load plugin

Available Tools

Tool	Description
pinchtab__navigate	Navigate to URL
pinchtab__snapshot	Get page structure
pinchtab__action	Click, type, press keys
pinchtab__text	Extract text content
pinchtab__screenshot	Capture screenshot
pinchtab__pdf	Generate PDF
pinchtab__evaluate	Run JavaScript
pinchtab__cookies_get	Get cookies
pinchtab__stealth_status	Check stealth mode

Usage in Claude Code

Use pinchtab to navigate to example.com and extract the main headlines.

Claude will:

1. Call pinchtab__navigate with URL
2. Wait 3 seconds
3. Call pinchtab__snapshot with filter
4. Extract headlines from result

---

Headless vs Headed

Aspect	Headless	Headed
Window	No visible UI	Chrome window visible
Speed	~20% faster	Slower (rendering overhead)
Memory	~50-80 MB	~100-150 MB
Use Case	CI/CD, scraping, batch	Debugging, visual QA
Interaction	API only	API + manual

# Headless for production
pinchtab instances create --mode=headless

# Headed for debugging
pinchtab instances create --mode=headed

---

Best Practices

DO

• ✅ Wait 3+ seconds after navigation
• ✅ Use text extraction over screenshots (token-efficient)
• ✅ Filter snapshots to reduce tokens
• ✅ Use profiles for persistent sessions
• ✅ Run headless in production
• ✅ Clean up instances after use
• ✅ Handle errors gracefully

DON'T

• ❌ Skip the 3-second wait
• ❌ Take screenshots for text extraction
• ❌ Parse full snapshots without filtering
• ❌ Use headed mode in CI/CD
• ❌ Leave instances running indefinitely
• ❌ Hardcode element refs (they change)

---

Troubleshooting

Only getting 1 node in snapshot

Cause: Accessibility tree not ready Fix: Increase wait time to 3+ seconds

pinchtab nav https://example.com
sleep 3  # Increase if needed
pinchtab snap

Timeouts

Cause: Page too slow or Chrome overloaded Fix: Increase sleep or use headless mode

# Increase wait
sleep 5

# Or use headless for faster rendering
pinchtab instances create --mode=headless

Element not found

Cause: Refs change between snapshots Fix: Re-snapshot before each action

# Get fresh refs before each action
REF=$(pinchtab snap -i | jq -r '.nodes[] | select(.name == "Submit") | .ref')
pinchtab click "$REF"

Connection refused

Cause: PinchTab server not running Fix: Start server first

pinchtab  # In separate terminal

---

References

• PinchTab Documentation
• PinchTab GitHub
• Agent Optimization Guide
• Common Patterns
• SMCP Plugin

---

Token-efficient browser automation for AI agents.