A LinkedIn AI agent with hands

A real LinkedIn AI agent session: Claude writes every invite and message over MCP, PoliteReach sends them from your own account under published caps.

Most demos of a LinkedIn AI agent stop at the draft. The model writes a sharp message, the founder nods, the video ends. Nobody shows the send. Sending is the part LinkedIn polices and the part chatbot wrappers quietly hand back to you. PoliteReach exists for the send.

The split is simple. Claude, connected over MCP, is the brain: it picks targets, writes every connection note and message, reads replies, drafts answers, and decides next steps. PoliteReach is the hands. A real headful Chromium browser runs in the cloud, signed into your own LinkedIn account, executing exactly what the brain approved. No extension, no injected code.

What an AI agent to send LinkedIn messages actually needs

A model that only drafts is a writing tool. An AI agent to send LinkedIn messages needs three things a chat window doesn't have: a real logged-in browser session, a way to detect invite acceptance, and limits that hold even when the model is wrong. PoliteReach supplies all three. (LinkedIn exposes no acceptance API, which is why the second one is harder than it sounds.)

The browser is headful Chromium in the cloud with one persistent device profile per account, behind an ISP-type IP we provide and check by ASN type, so a datacenter range never carries your account. It looks like your laptop. Acceptance is inferred by diffing: PoliteReach compares your current connections against still-pending invites and records the overlap as accepted. The caps live in the executor. A prompt can't reach them. The tools are plain MCP; Claude is the default brain, but any MCP client or custom agent can drive them.

What one agent session looks like, tool by tool

Each step in the run below is the actual MCP tool the assistant calls, not a metaphor.

Claude ingests 20 prospects, sequences attached

Claude calls the ingest tool with 20 profile URLs. Each prospect carries a personalized connection note plus prewritten Day 1, Day 3, Day 7, and Day 14 messages, all drafted by Claude and reviewed by you before anything queues. None of it changes later. Nothing is generated at send time; if a message exists, you've already read it.

PoliteReach drips the invites under caps

No tool call here; the executor takes over. Invites go out at the account's pace, 18 to 24 a day on a verified one, jittered, inside office hours, each note typed key by key. The note rides along from a Premium account; on a free one LinkedIn drops it. The cap doesn't negotiate. Claude queued 20 prospects; it could queue 500 and the daily number wouldn't move.

PoliteReach spots acceptances by diffing connections

No tool call here either. PoliteReach checks on its own about every 90 minutes through the day. Two days in, it diffs the connection list against pending invites and finds three matches, so three acceptances get recorded and three Day 1 messages get scheduled off each acceptance time. The other seventeen keep waiting. If you have switched the cleanup on, anything still pending after 60 days gets withdrawn for you; the window is configurable down to a floor of 7.

A reply halts the sequence before Claude sees it

One accepted prospect answers the Day 1 message. The instant that reply lands, every remaining step for that prospect stops, permanently. Claude then calls the conversation-list tool, spots the new reply, and pulls the full thread with real timestamps through the read tool. Now it drafts.

You approve the reply before it ships

Claude writes an answer in the thread's context and puts it in front of you. You edit or approve; the send tool ships it through the same cloud browser. Messages to people you're already connected with aren't invite-capped, so a live conversation moves at conversation speed. Deals aren't rate-limited.

Claude posts to your feed between runs

Before signing off, Claude calls the post tool with a short take your prospects would actually read; the tool accepts up to 3,000 characters and an optional image. Posting matters. An account that only sends invites reads as a bot; one that also talks reads as a person.

Every word PoliteReach sends existed, and was approved, before the first invite went out.

You can use AI to send LinkedIn messages and still approve every word

The division of labor is strict. You own strategy and approval: who to target, what the offer is, what each draft says before it goes anywhere. Claude proposes, PoliteReach executes, and the full action log records every step either of them took. If something feels wrong, revoke your LinkedIn session cookie and everything stops that instant, because the session is yours, not ours. That's real control.

The brain is swappable too. Change the model, the prompt, or the whole strategy any day; the executor doesn't care who wrote the words. One subscription per team, priced per LinkedIn account, and no per-contact send credits.

The caps the agent can't talk its way past

Limits are enforced in the execution layer, not in the prompt. Claude can't raise them, and neither can a prompt injection. The published set:

• 5-10 invites a day for a new or unverified account, 18-24 a day once verified, 25-30 a day on Premium, jittered
• A rolling 7-day allowance per account to match: 63-70 a week new, 90-100 verified, 135-150 Premium. LinkedIn publishes no invitation limit, so these numbers are ours, and an account that hits LinkedIn's wall anyway drops a tier for eight days
• Invites and automated follow-ups sent during your sending hours only
• Optional 4-week warm-up ramp
• Connection notes typed key by key; messages pasted and read back line by line before Send
• Per-account circuit breaker that freezes jobs on a security checkpoint
• Optional cleanup of invites pending more than 60 days, off until you switch it on (configurable, floor of 7)

Messages to existing connections are deliberately uncapped, because invites are the action LinkedIn punishes and conversations are the point. A checkpoint on one account never pauses the others.

The caps live in the code that does the sending, so no prompt, model swap, or bad day can move them.

What a LinkedIn automation AI agent should cost

PoliteReach costs $59 a month per LinkedIn account, so the honest comparison is stack cost, meaning what you'd spend to run the whole thing: that seat plus a Claude subscription at $20 to $200 a month (claude.com/pricing). The proxy and the server come with the seat. Here's the field.

Two caveats. The 11x and Artisan numbers are sales-quoted or reported rather than list prices, and the AI SDR platforms bundle lead databases plus email sending that this stack doesn't include. Both matter. Computed against a $159 stack, the gaps run about 6x for AiSDR's $900 Explore tier, 6x to 16x for Artisan, 11x for Regie.ai, and roughly 31x for 11x. Seat tools cost about what the PoliteReach seat does. The difference there is the brain: Dripify schedules templates, while an agent reads the thread and decides the next move.

Invites, messages and posts on your own account

LinkedIn invites, messages, and feed posts on your own account, driven by the AI you already pay for. That's the surface. If it's the half of outreach you've been doing by hand, put your AI on it. The first 14 days are free.