Agent-Skills.md

Agent Skills: LinkedIn CDP Automation

LinkedIn CDP: Input-only automation (mouse, keyboard, screenshots). Zero DOM access.

UncategorizedID: aaaaqwq/claude-code-skills/linkedin-cdp

Repository

aaaaqwq/claude-code-skills
aAAaqwqLicense: MIT
7317

Install this agent skill to your local

pnpm dlx add-skill https://github.com/aAAaqwq/AGI-Super-Team/tree/HEAD/skills/linkedin-cdp

Skill Files

Browse the full folder contents for linkedin-cdp.

Download Skill

Loading file tree…

skills/linkedin-cdp/SKILL.md

Skill Metadata

Name
linkedin-cdp
Description
LinkedIn CDP: Input-only automation (mouse, keyboard, screenshots). Zero DOM access.

LinkedIn CDP Automation

LinkedIn automation via Chrome DevTools Protocol. Input-only (mouse/keyboard/screenshots). Zero DOM access.

When to use

  • "collect messages from linkedin" -> LinkedInMessages
  • "who messaged me on linkedin" -> LinkedInMessages
  • "write on linkedin" -> LinkedInMessages.send_message()
  • "find on linkedin" -> LinkedInSearch
  • "view profile" -> LinkedInProfile
  • "send connection request" -> LinkedInConnect
  • "show connection requests" -> LinkedInConnect.screenshot_invitations()

Dependencies

  • External: Chrome, pip install websocket-client requests
  • Chrome must run with --remote-debugging-port=9222 (separate instance)

Paths

| What | Path | |------|------| | Script | $HOME/linkedin-cdp/linkedin_cdp.py | | Modules | $HOME/linkedin-cdp/linkedin_*.py | | Rate limiter | $HOME/linkedin-cdp/rate_limiter.py | | Screenshots | /tmp/li_screenshots/shot_*.jpg |

How to execute

Step 0: Chrome Launch

IMPORTANT: Use the binary path directly, NOT open -a 'Google Chrome'. open -a on macOS opens a tab in existing Chrome instead of a separate instance.

/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 \
  '--remote-allow-origins=*' \
  --user-data-dir="$HOME/chrome-debug-profile" \
  "https://www.linkedin.com" > /dev/null 2>&1 &

Run in background (run_in_background: true), then verify:

sleep 3 && curl -s "http://localhost:9222/json/version"
  • --user-data-dir MUST differ from user's main Chrome profile
  • $HOME/chrome-debug-profile persists login session between runs
  • Don't connect CDP while user logs in. Wait for login to complete.

Step 1: Calibration (REQUIRED on first run per session)

Every session may have different window size / DPR. On first run, take a calibration screenshot and determine the coordinate mapping:

import sys, subprocess, time
sys.path.insert(0, '$HOME/linkedin-cdp')
from linkedin_cdp import LinkedInBot

bot = LinkedInBot()
bot.connect()

# Take calibration screenshot — returns file path
path = bot.take_screenshot()
print(path)  # /tmp/li_screenshots/shot_0001.jpg

# Get image dimensions to determine DPR
result = subprocess.run(['sips', '-g', 'pixelWidth', '-g', 'pixelHeight', path],
                       capture_output=True, text=True)
print(result.stdout)
bot.close()

Then read the calibration screenshot with Read tool. Calculate:

  • DPR = image_width / expected_viewport_width (typically 2 on Retina Mac)
  • CSS coordinates = image_pixel_coordinates / DPR

Store the DPR for all subsequent coordinate calculations in this session.

Step 2: Coordinate System

All click_at(), _click() calls use CSS coordinates, not image pixels.

Formula: CSS_coord = image_pixel_coord / DPR

Typical Retina Mac (DPR=2, viewport ~1531x801):

  • Screenshot: 3062x1602 pixels
  • To click where you see something at image pixel (600, 400): click CSS (300, 200)

Step 3: Use modules

import sys
sys.path.insert(0, '$HOME/linkedin-cdp')
from linkedin_cdp import LinkedInBot

bot = LinkedInBot()
bot.connect()

# All actions use CSS coordinates
bot.click_at(x, y)        # click + screenshot
bot.type_text("text")      # type with human-like delays
bot.scroll_wheel(delta_y=500)  # scroll (keyword arg only!)
bot.take_screenshot()      # returns file path (JPEG)
bot.navigate_to(url)       # navigate + auto reconnect
bot.reconnect_to_tab()     # reconnect WebSocket after page change
bot.close()

Known Fixed Coordinates (DPR=2, viewport ~1531x801)

These modal/dialog coordinates are consistent across sessions at the same viewport size. Recalibrate if window size changes.

Connection Request Modal

After clicking "Connect" on a profile:

| Element | CSS (x, y) | Notes | |---------|-----------|-------| | "Add a note" button | (751, 239) | White/outline button | | "Send without a note" button | (911, 239) | Blue button | | Note text field (click to focus) | (752, 259) | After clicking "Add a note" | | "Send" button (with note) | (968, 393) | Blue, active after typing | | "Cancel" button | (889, 393) | |

Profile Page

| Element | CSS (x, y) | Notes | |---------|-----------|-------| | "Connect" button | (~270, 467-499) | y varies by profile layout (banner height, etc.) | | "Message" button | (~383, 467-499) | Next to Connect |

Messaging Page

/messaging/ — conversation list on left, active thread on right.

| Element | CSS (x, y) | Notes | |---------|-----------|-------| | "Write a message..." text field | (715, 604) | Click to focus, then type_text() | | Send button | (885, 724) | Active (blue) only after typing |

Invitation Manager

/mynetwork/invitation-manager/ — lists received/sent invitations.

| Element | CSS (x, y) | Notes | |---------|-----------|-------| | "Accept" button (first invite) | (~920, 274) | Approximate — verify via screenshot | | "Ignore" button (first invite) | (~838, 274) | Next to Accept | | "Received" / "Sent" tabs | (~265, 142) / (~350, 142) | Top of page |

Search Results

| Element | CSS (x, y) | Notes | |---------|-----------|-------| | First result name | (~305, 155) | Approximate — always verify via screenshot | | Second result name | (~305, 280) | Approximate — positions shift with ads/banners |

Architecture

LinkedInBot (linkedin_cdp.py)  -- base class
   |-- CDP core: connect(), _send(), close(), reconnect_to_tab()
   |-- Mouse: _bezier(), _human_path(), _move_to(), _click(), _maybe_fake_hover()
   |-- Keyboard: type_text(), press_key()
   |-- Scroll: scroll_wheel()
   |-- Screenshot: take_screenshot() -> path, take_screenshot_base64(), save_screenshot()
   |-- Navigation: navigate_to(), wait_for_page()
   |-- Convenience: click_at(x,y) -> screenshot, scroll_and_screenshot()
   |
   +-- LinkedInMessages   (linkedin_messages.py)
   +-- LinkedInSearch      (linkedin_search.py)
   +-- LinkedInProfile     (linkedin_profile.py)
   +-- LinkedInConnect     (linkedin_connect.py)

Modules

| File | Class | Key Methods | |------|-------|-------------| | linkedin_cdp.py | LinkedInBot | click_at(), type_text(), scroll_wheel(), take_screenshot() -> path, take_screenshot_base64(), navigate_to() | | linkedin_messages.py | LinkedInMessages | screenshot_conversations(), open_conversation(), send_message(), collect_screenshots() | | linkedin_search.py | LinkedInSearch | search_people(), search_companies(), next_page() | | linkedin_profile.py | LinkedInProfile | view_profile(), screenshot_full_profile(), scroll_to_section() | | linkedin_connect.py | LinkedInConnect | view_profile(), send_connection_note(), screenshot_invitations(), accept_invitation() | | rate_limiter.py | RateLimiter | can_search(), can_view_profile(), wait_if_needed() |

Examples

Search → Profile → Connect with Note (full flow)

import sys, time
sys.path.insert(0, '$HOME/linkedin-cdp')
from linkedin_cdp import LinkedInBot

bot = LinkedInBot()
bot.connect()

# 1. Search
bot.navigate_to('https://www.linkedin.com/search/results/people/?keywords=Name%20Company')
time.sleep(4)
bot.reconnect_to_tab()
time.sleep(2)
# take_screenshot() returns file path — read with Read tool to find coordinates

# 2. Click profile (coordinates from screenshot / DPR)
path = bot.click_at(305, 155)  # first result — verify from screenshot!
time.sleep(4)
bot.reconnect_to_tab()
# Read path with Read tool to verify correct profile

# 3. Click Connect (~270, 467-499 — verify from screenshot)
bot.click_at(270, 499)
time.sleep(2)
# Modal appears

# 4. Click "Add a note" (fixed coordinate)
bot.click_at(751, 239)
time.sleep(1.5)

# 5. Click text field + type note
bot._click(752, 259)
time.sleep(0.5)
bot.type_text("Hi Name, personalized note here...")
time.sleep(1)

# 6. Click Send (fixed coordinate)
bot.click_at(968, 393)
# Done — verify "Pending" status + "Invitation sent" toast

bot.close()

Read Messages

from linkedin_messages import LinkedInMessages
lm = LinkedInMessages()
lm.connect()

path = lm.screenshot_conversations()
# path = '/tmp/li_screenshots/shot_NNNN.jpg' — read with Read tool

path = lm.open_conversation(200, 350)  # coordinates from screenshot
# Read path with Read tool

lm.close()

View Profile (manual scroll)

import time
from linkedin_profile import LinkedInProfile

prof = LinkedInProfile()
prof.connect()
prof.navigate_to('https://linkedin.com/in/username')
time.sleep(4)
prof.reconnect_to_tab()

paths = [prof.take_screenshot()]
for _ in range(4):
    prof.scroll_wheel(delta_y=500)
    time.sleep(1.5)
    paths.append(prof.take_screenshot())
# paths = ['/tmp/li_screenshots/shot_0001.jpg', ...] — read each with Read tool
prof.close()

Troubleshooting

| Problem | Solution | |---------|----------| | WebSocketConnectionClosedException after navigate | Call reconnect_to_tab() — page change breaks WebSocket | | Click misses target | Verify DPR: image_pixels / DPR = CSS coords. Re-run calibration. | | screenshot_full_profile() crashes | Use manual navigate + reconnect + scroll loop (see View Profile example) | | scroll_wheel() TypeError | Use keyword arg only: scroll_wheel(delta_y=500), NOT positional args | | Chrome opens tab in existing browser | Use binary path directly, NOT open -a 'Google Chrome' | | CDP port conflict | Change --remote-debugging-port=9223 and update CDP_PORT in linkedin_cdp.py | | Modal coordinates wrong | Window resized? Re-run calibration. Fixed coords assume viewport ~1531x801. |

Practical Tips

  1. LinkedIn URLs from web research are often wrong. Always search by name + company.
  2. scroll_wheel() takes delta_y keyword arg only. Not positional.
  3. After any navigate_to() or page-changing click, always reconnect_to_tab().
  4. Always sys.path.insert(0, '$HOME/linkedin-cdp') before imports.
  5. Screenshots auto-save to /tmp/li_screenshots/shot_*.jpg — read them with Read tool. Old files auto-cleaned (keeps last 50).
  6. One module instance at a time. Close previous before opening new.
  7. Calibrate DPR on first run. Don't assume DPR=2 — verify.
  8. No em-dashes in messages. Never use long dashes (--) in connection notes or messages. People recognize it as AI-generated text. Use commas, periods, or short sentences instead.

Rate Limits (recommended daily)

| Action | Conservative | Moderate | |--------|-------------|----------| | Profile views | 50 | 100 | | Searches | 20 | 50 | | Connection requests | 15 | 25 | | Messages sent | 30 | 50 |

Security Model

Zero DOM access — NEVER Runtime.evaluate, querySelector, innerText. Screenshot-based reading — Page.captureScreenshot (JPEG, quality 80) saved to files, not base64 in memory. Human-like mouse — unique Bezier curves, tremor, overshoot, micro-pauses, speed variation. Real Chrome — not headless. Normal fingerprint. Rate limiting — built-in daily caps. Human-in-the-loop — Claude reads screenshots, adds natural irregularity.

NEVER do

  • NEVER use Runtime.evaluate or CDP Runtime domain — #1 way bots get caught
  • NEVER bypass rate limits — even if asked to "go faster"
  • NEVER run headless — always visible Chrome window
  • NEVER automate login — user logs in manually

Post-outreach CRM logging (REQUIRED)

After sending a connection request or message, ALWAYS log to CRM:

  1. Company in companies.csv (if new)
  2. Person in people.csv (if new)
  3. Lead in leads.csv (stage=new, source=linkedin, source_direction=outbound)
  4. Activity in activities.csv with the exact message text in notes field:
    • type: outreach
    • channel: linkedin
    • direction: outbound
    • subject: "LinkedIn connection request with note" (or "LinkedIn message")
    • notes: the full text of the message sent

Do NOT skip step 4. The message text must be preserved in CRM for follow-up context.

Related skills

  • add-lead — add found contacts to CRM
  • update-lead — update lead status after outreach
  • log-activity — log activity to activities.csv
  • email-send-bulk — follow up via email after LinkedIn connect