How to Run LinkedIn Outreach With Claude

Connect Claude to your outreach workspace over MCP and you can run LinkedIn in conversation: create a campaign, enrol prospects, start it, pause it, publish a post, and ask why yesterday was quiet. Claude never sends an invite or a message itself and can never raise a limit — the scheduler paces every action inside per-account safety caps.

Key takeaways

• One command connects Claude, then a browser login. No code.
• Claude can create, start, pause, resume and enrol LinkedIn campaigns, publish and manage posts, and run job postings.
• Claude cannot send. Launching hands the campaign to the scheduler, which decides what goes out and when.
• Claude cannot raise a cap — not the daily invite allowance, not the ramp, not the per-account pace. No such tool exists.
• Connecting the LinkedIn account stays in the app, deliberately.

Why does the send boundary matter so much on LinkedIn?

Because the downside is asymmetric. A cold email mailbox that gets throttled recovers. A restricted LinkedIn account frequently does not, and everything built on it — the connections, the conversation history, the profile itself — goes with it.

That is why the execution split is not a limitation to work around but the reason the setup is safe to use at all. An AI agent works at agent speed: it will happily enrol 400 prospects in one turn. If that instruction reached a live send path, you would have a burst of invites from one account inside a minute, which is precisely the pattern that gets accounts restricted. Instead the instruction writes campaign state, and a scheduler that knows the account’s daily allowance, its ramp position and its rest windows releases actions gradually.

Account safety wins over speed, over throughput, and over what you asked for. If a configuration conflicts with a safety limit, the step waits.

How do you connect Claude?

Per Anthropic’s documentation, Claude Code adds a remote MCP server with:

claude mcp add --transport http warmysender https://warmysender.com/mcp

HTTP is the recommended transport. For servers that use OAuth, you add the server first and then run /mcp inside Claude to complete a browser login — Anthropic’s docs show the connection status right there (✔ Connected, ! Needs authentication), and tokens refresh automatically. From v2.1.186 you can authenticate from the shell with claude mcp login warmysender. Add --scope user if you want the connection available in every project rather than the current one.

Alternatively, authenticate with a scoped key you paste once. Either route resolves to the same permissions.

MCP itself is the open standard that makes this work — the protocol documentation describes it as a standardised way for AI applications to reach external systems, which is why the same server also works with ChatGPT, Cursor, Codex, OpenClaw, Hermes Agent, and any other client that speaks it.

What can Claude do with LinkedIn once connected?

Build and run campaigns. Create a LinkedIn campaign, start it, pause it, resume it, and enrol prospects into it. Claude writes the campaign and hands it over; it never issues an invite or a message directly.

Publish content. Create, publish and manage LinkedIn posts — useful when outreach and presence are meant to move together.

Run hiring workflows. Create job postings, publish them, close them, and read applicants.

See the account honestly. List accounts and read account health, including whether the account is currently in a protective rest during which nothing will send. Read a campaign’s throttle picture and its stats. Ask “why did this campaign send nothing today?” and get a plain-language answer rather than an empty dashboard.

That last capability is the one that changes day-to-day work. Most “my campaign is broken” moments are a paced first send, a start date in the future, a rest window, or an account tier that has not ramped yet — all things Claude can read and explain in a sentence.

Work across channels in one conversation. The same connection covers cold email and Instagram, so “pause the LinkedIn campaign and enrol the non-responders into the email sequence” is one request.

What will Claude refuse to do?

• Send an invite, message or InMail directly. There is no tool that does it. Every action goes through the scheduler.
• Raise any limit. Not the daily invite allowance, not the messages cap, not the ramp step for a new account.
• Connect or disconnect a LinkedIn account. That stays in the app — exposing it would let a compromised agent drain seats or trip account checkpoints.
• Delete anything. Pause and archive cover the real cases and are reversible.

These are absences, not refusals the model chooses. A more persuasive prompt does not unlock them.

What does a good first session look like?

Something like this, in order:

1. Ask for the state. “List my LinkedIn accounts and their health.” Confirm the account is connected, not resting, and where it sits in its ramp.
2. Check the caps before planning volume. “What does the throttle look like on this account?” A new account starts deliberately slow, and planning around its real allowance beats discovering it later.
3. Create the campaign, don’t start it. Have Claude write the steps and messages, then read them back yourself. Personalisation errors are far cheaper to catch before launch.
4. Enrol a small batch first. Twenty prospects, not four hundred. Enrolment is reversible, but a small first cohort tells you whether your placeholders resolve properly.
5. Start it and leave it alone. The first day will look slow because it is supposed to.
6. Come back and ask why. “Why did this campaign send nothing in the last day?” is the fastest diagnostic you have.

How WarmySender handles this

WarmySender’s MCP server is what Claude connects to, and every LinkedIn tool routes through the same scheduler the app uses — so a cap enforced for a person clicking a button is enforced identically for Claude calling a tool. New accounts ramp gradually rather than starting at full allowance, campaigns sharing one account share its daily limit rather than each getting their own, and an account in a protective rest simply defers instead of pushing through.

The LinkedIn safety FAQ covers the safety model, LinkedIn sending limits and account tiers documents the caps themselves, and how WarmySender protects your LinkedIn account explains what happens when a limit binds. If you want the wider picture of agents driving LinkedIn alongside leads, start with the agents, LinkedIn and leads walkthrough.

Frequently asked questions

Does Claude need my LinkedIn password?

No. Connecting the LinkedIn account happens in the app and Claude never sees those credentials. Claude authenticates to your workspace separately — with a scoped key or a browser sign-in — and works through tools rather than through your account directly.

What happens if I ask Claude to send 500 invites today?

It will enrol the prospects, and then nothing dramatic happens. The scheduler releases invites at the account’s allowance, in its ramp position, inside its sending window. The campaign simply runs for as long as the safe rate requires. If you genuinely need more throughput, the answer is more connected accounts, never a higher per-account number.

Can Claude reply to LinkedIn messages for me?

Claude can read campaign state and stats, but sending — including replies — is not an agent capability. Conversations happen in the app. The design keeps the model out of the live send path entirely.

Why is my campaign sending fewer invites than the limit allows?

Usually the account ramp: a newly connected account starts well below its eventual allowance and grows over weeks. Other common causes are multiple campaigns sharing one account’s daily limit, a protective rest window, or a start date in the future. Ask Claude why the campaign sent nothing — that explanation is one of its tools.

Can other agents do the same thing?

Yes. The capabilities live in the server, not the client, so ChatGPT, Cursor, Codex, OpenClaw and Hermes Agent — and anything else that speaks MCP — see the same LinkedIn tools with the same limits. What differs is how the connector is added.