LinkedIn Automation¶

Automate your LinkedIn connection workflow directly from Claude Code. Accept pending invitations, send outreach messages, and export people search results to a Google Sheet — all by typing a slash command.

No coding required after setup.

What this does¶

Command	What it does
`/invites`	List all your pending connection invitations
`/accept-one`	Accept the next pending invitation and send your message template
`/accept-many N`	Accept the next N invitations and message each one
`/message-person <url>`	Send a message to one connection
`/message-connections N`	Message N existing connections who haven't been messaged yet
`/search-to-sheet <url>`	Extract LinkedIn people search results and save to CSV / Google Sheet
`/linkedin-search-capture "<keywords>"`	Search by keyword / location / degree and capture results to a JSON file
`/linkedin-find-person "Name \\| Org \\| Country"`	Find someone's profile URL by name
`/connect-from-sheet <sheet>`	Send connection requests to everyone in a spreadsheet

Prerequisites¶

Before using any of these commands, you need:

Claude Code installed — get it here
Claude in Chrome browser extension — install it from the Chrome Web Store, then grant it access to linkedin.com
Chrome open and logged into LinkedIn
A Google Sheet (only needed for /search-to-sheet)

Setup¶

Step 1 — Create your message template¶

This is the message that gets sent automatically after you accept an invitation.

Copy the example file:

cp automation/linkedin-automation/setup/accept-invite.txt.example texts/accept-invite.txt

Open texts/accept-invite.txt in any text editor and write your message. Keep it under ~500 characters.

Step 2 — Set up Google Sheets (only for `/search-to-sheet`)¶

Skip this step if you don't plan to use /search-to-sheet.

Open a new Google Sheet
Click Extensions → Apps Script
Delete the default code and paste the contents of automation/linkedin-automation/scripts/append-rows.gs
Click Deploy → New deployment → Web app
Execute as: Me
Who has access: Anyone
Click Deploy and copy the Web App URL

Copy the config example file and fill in your URL:

cp automation/linkedin-automation/setup/config.json.example config.json

Open config.json and replace YOUR_GOOGLE_SHEETS_WEB_APP_URL_HERE with your URL

Step 3 — Open this folder in Claude Code¶

Open the METIS-pub folder in Claude Code. The slash commands will be available automatically.

Using the commands¶

Check your pending invitations:

/invites

Accept one invitation (and send your message):

/accept-one

Accept 5 invitations at once:

/accept-many 5

Message 10 existing connections:

/message-connections 10

This skips anyone you've already messaged (tracked in messaged-log.txt).

Export a LinkedIn search to Google Sheet:

/search-to-sheet https://www.linkedin.com/search/results/people/?keywords=...

Capture a filtered search to a JSON file:

/linkedin-search-capture "AI for good" --degree 2nd --pages 10
/linkedin-search-capture --location "Philippines" --degree 1st --pages 2

Filters by keyword, location and connection degree. Location names are resolved to LinkedIn's internal location IDs and cached in geo-ids.json, so each place costs that lookup only once.

Pacing — why runs take a while¶

These commands deliberately go slowly. Every workflow that loops over people or pages waits a randomised amount of time between steps, drawn from an exponential distribution so the rhythm resembles a person browsing rather than a script. Roughly:

Action	typical gap
Next page of search results	~12 s
Visiting a profile	~45 s
Accepting an invitation	~30 s
Sending a message	~90 s
Sending a connection request	~2 min

Every 8–15 actions there is also a longer break of 2–10 minutes.

So /connect-from-sheet with 20 rows is a 45–60 minute job, not a few seconds. Each command tells you the expected duration before it starts, and announces every long break with its length and where it has got to — a silent pause is never a hang. Short per-item delays aren't announced, because the progress lines already show the run is alive.

Everything is written as it goes (the sheet, the log, the output file), so a run you stop — or one that stops itself — picks up cleanly where it left off.

Rate limiting is not an error¶

If you push LinkedIn too hard it doesn't return an error. It quietly serves a reduced profile page: the right name, headline and location, but no About text, no follower count, no company or school. Everything looks like it worked while most of the data is missing.

The commands detect this — a missing About together with a missing follower count means throttled, since a genuine profile with no About still shows its follower count. When that happens the run stops and tells you how far it got, rather than recording empty fields as if they were real. Give it 15–30 minutes before trying again.

Profile visits are visible¶

Reading a profile is not anonymous — the person is shown that you viewed them. A bulk enrichment run over 100 people means 100 people see that. Worth deciding deliberately before starting one.

How it works¶

Each command is a markdown file in .claude/commands/. When you type /accept-one, Claude Code reads that file and follows the step-by-step instructions — navigating your Chrome browser, clicking buttons, reading page content, and sending messages, all using the Claude in Chrome extension.

You don't need to understand any of this to use it. Just type the command.

Files created automatically¶

texts/accept-invite.txt — your message template (you create this in setup)
config.json — your Google Sheets URL (you create this in setup)
messaged-log.txt — auto-created, tracks who you've already messaged so you don't duplicate
output/search-results.csv — auto-created when you run /search-to-sheet
output/linkedin-search-*.json — auto-created when you run /linkedin-search-capture

These files are excluded from git (see .gitignore) so your personal data stays local.

One file that is committed: automation/linkedin-automation/geo-ids.json, the cache mapping place names to LinkedIn's internal location IDs. It holds no personal data, and sharing it means nobody has to redo the lookup for a place someone has already resolved.

Troubleshooting¶

Command not working / Claude seems confused

Make sure the Claude in Chrome extension is active on the LinkedIn tab
Check that Chrome is open and you're logged into LinkedIn
Try closing and reopening the LinkedIn tab, then run the command again

LinkedIn page structure changed

LinkedIn periodically updates its DOM. If a command fails with "element not found" errors, open a GitHub issue in this repo describing what happened.

A run stopped saying it was rate limited

This is working as intended. LinkedIn started serving reduced pages, and the command stopped rather than saving half-empty records. Wait 15–30 minutes and run it again — it resumes from where it stopped.
If it happens repeatedly, leave it for the day. Nothing is lost; everything already captured has been written.

A command seems to be doing nothing

Check whether it announced a break (Taking a 6 min break…). Pauses of 2–10 minutes are normal and deliberate — see Pacing.
Runs are meant to take minutes to hours. If you need a quick result, ask for fewer items rather than expecting it to go faster.

Messages going to wrong people / duplicates

Check messaged-log.txt — it lists everyone you've already messaged
The /message-connections command reads this file before sending anything