What AI does on LinkedIn, and what each way in hands LinkedIn
An assistant cannot open LinkedIn. Give Claude or ChatGPT a profile URL and it will either tell you it cannot open the page or invent what is on it — LinkedIn blocks assistant crawlers, and no self-serve permission in LinkedIn's own API returns another member's profile. Everything in this section is a different answer to that. You paste the page in, which costs nothing and exposes nothing. You use the 29 AI features LinkedIn ships inside the product. You buy one of 42 third-party tools, 18 of which act under your name. Or you connect one of 13 MCP servers, none of them LinkedIn's. They differ by what they hand LinkedIn, and that is the part this site can date and source.
What an assistant genuinely contributes, and what stays manual
The 8 areas of work below, in the order the work is done rather than by search volume. Read the right-hand column first. Every published list of LinkedIn AI use cases stops at the middle one, which is how people end up holding a plan they cannot execute and blaming the model for it.
| What you arrived asking | What the assistant does | What it cannot do, and why | Start here |
|---|---|---|---|
| Who should I actually be talking to? | Scores a pasted or exported list against a written ICP and hands it back ranked, with the reason for each placement and a line for the ones it could not judge. | It cannot go and find the people. LinkedIn has no public search API and blocks assistant crawlers, so the list has to arrive from you. | Prospecting |
| How do I find them without guessing at the filters? | Turns an ICP paragraph into a filter set and the Boolean strings to go with it, and marks which filters exist only on Advanced. | It cannot click the filters. The thinking takes ten seconds and the last three minutes in the interface are yours. | Sales Navigator |
| What do I send, and how often? | Drafts to a stated length against a checkable reference, and sizes the week against the invitation ceiling rather than past it. | It cannot send, and it cannot confirm that the thing it referenced is true. Read the draft before a person does. | Outreach |
| What should my profile say? | Rewrites a headline and an About section against the terms a recruiter actually searches, using your own material rather than inventing achievements. | It cannot see your profile as it stands, and it has no view of how you rank for anything. Paste the current text in. | Profile |
| What do I post, in my own voice? | Turns raw material — a call transcript, a document, a half-formed argument — into posts that keep the sentence rhythm of samples you supply. | It has no access to your analytics and no idea what your audience engaged with. Nothing here predicts reach. | Content |
| How do I fill this role with 30 InMails? | Turns a job description into a screening rubric, then applies that rubric to pasted profiles with a stated confidence for each candidate. | It cannot open a profile, and it should not be the thing that rejects a person. The rubric orders your reading; it does not replace it. | Recruiting |
| How do I get a reply from a hiring manager? | Matches your experience against a pasted posting, names the gaps plainly, and drafts the note the posting is actually asking for. | It cannot apply for you, and it cannot tell you whether the role was filled internally three weeks ago. | Job search |
| How do I run this for six clients at once? | Keeps each client’s ICP, voice and permitted claims in its own context file, so six clients are six folders instead of six chats you re-explain. | It cannot keep them apart if you do not. One conversation holding two clients’ files will blend them, confidently. | Agencies |
The right-hand column is mostly one fact repeated. LinkedIn does not let an assistant read it, and it does not let one act on it either — the ceilings that decide how much of the middle column you can actually use are on the limits page, and they apply on every plan including the most expensive one.
Three of these rows are one job in sequence for anyone selling — prospecting, search and outreach — and the sales hub runs them in that order against the weekly invitation ceiling rather than one row at a time.
The routes differ by what they hand LinkedIn, not by how clever the model is
What separates these is not the model. It is the credential, and the credential decides two things nothing else does: how much of LinkedIn the thing can reach, and who carries the consequence when LinkedIn objects. Read the second and third rows before the rest — they are the ones that cannot be undone by choosing a better product inside the same column.
| You paste it in50 free resources | LinkedIn’s own AI29 features | A third-party tool42 researched | An MCP server13 community-built | |
|---|---|---|---|---|
| What it hands LinkedIn | Nothing. The page goes from your clipboard to the model. | Nothing. LinkedIn is the vendor. | A signed-in session, a browser extension, or a vendor API key. | Whatever the server authenticates with. |
| A LinkedIn account sits behind it | No | Yours, inside the product | 29 of 42 | 9 of 13 |
| It can act on LinkedIn under your name | No | It drafts; you send | 18 of 42 | 8 of 13 |
| What it costs | Free | 5 on any account; the cheapest plan that adds more is $39.99 a month | 35 of 42 publish a figure at all | Free to run. The LinkedIn seat it needs is not. |
| Where the work happens | In your assistant | Inside LinkedIn | On the vendor’s servers, or in your browser | In your assistant, through the server |
Enforcement lands on the member account, not on the vendor. The Prohibited Software and Extensions policy names browser extensions and third-party automation services and states that LinkedIn may restrict or permanently restrict accounts that use them. The vendor keeps its subscription revenue either way. www.linkedin.com ↗. That clause is why the exposure column exists and why nothing in this section tells anyone to automate an account. Figures read from each registry on its own date: resources 18 September 2026, LinkedIn's own features 20 September 2026, tools and servers 20 September 2026.
The section, in the order you would actually use it
Settle what you already own before you buy, and what a route costs your account before you compare features inside it. Each card is a separate registry with its own dated records, and none of them ranks anything. If you do not know which card is yours, the stack picker asks 7 questions and returns the stack those answers justify, priced, or says there is nothing to buy.
What LinkedIn ships itself
Settle this before buying anything. LinkedIn ships 29 AI features across 6 of its products, 5 of them on every account, and several of the jobs below are already solved inside the product on a plan you may be paying for. 10 of the 29 have no price that can be looked up anywhere, and that page says which and why instead of leaving the column blank.
29 features, read 20 September 2026“I want to paste something that works.”Prompts and context files
28 prompts, each declaring a role, decision logic, an output schema and a refusal rule for facts it cannot check. The 15 .md context files are the part almost nobody publishes: attach ICP.md, PRODUCT.md and LINKEDIN-LIMITS.md to a Project once and every later request inherits them, instead of you retyping your business every session.
Free, no account“Which one do I run first?”Workflows
The answer to the standing criticism of prompt libraries, which is that they are a pile. 4 sequences run the resources in order and every step names what it hands the next one — the outbound one runs 8 steps from an ICP to a booked meeting, and each step is a resource you can also use on its own.
4 workflows“Which tool should I buy, and what does it do to my account?”Third-party tools
42 products, each read from the vendor’s own site and pricing page on 20 September 2026 and sorted by how it reaches LinkedIn rather than by feature count. 9 of them do not touch LinkedIn at all — they cover email and the phone, and are bought every week by people who believe otherwise. 18 act under your name.
42 tools, 275 sources“Can an agent do this instead of me pasting?”MCP servers and agents
13 LinkedIn MCP servers, none of them LinkedIn’s. They divide by what they authenticate with — nothing at all, a vendor API key, LinkedIn’s own OAuth, a session cookie, a driven browser — and that, rather than the tool count in the README, decides what a server can reach and who carries the consequence. Each profile carries the project’s own config block for Claude Desktop, Claude Code, Cursor and VS Code.
13 servers, 4 need no LinkedIn account“I want the whole folder, not one prompt.”Agent packs
7 packs, each a folder rather than a snippet: CLAUDE.md at the root, the context files it names, a PROMPTS/ directory. Unzip into a Claude Project, a Cursor workspace or an agent repo and the assistant starts the first conversation already knowing your ICP, your claims and the platform ceilings.
7 packsUnder those sit the parts that are useful on their own. The 15 context files have their own index, because they work without any prompt at all. The prompt builder assembles one from a goal, a role and an output, with the limits already in it, and runs in your browser with no model call. Prompt check reads a prompt you already use and names the structural faults it has — no role, no output format, no guard against invented facts. And the Starter Pack is the assembled subset, with the licence. The MCP card splits one way the others do not: the client you are in decides how much of that directory you can reach at all, which is why Claude, Cursor and ChatGPT have a page each rather than a row in a table.
Paste the prompt. Get a structured, evidence-checked answer.
Not a paragraph promising quality — the actual worked example from LinkedIn Prospect Research, unedited. Every prompt in the library returns something with this shape: a stated hypothesis, a confidence rating, and an explicit line for what it could not verify.
Research Dana Whitfield, Group Finance Director at Northbeam Logistics, before I message her. Here's her profile and her latest post.
Value hypothesis — Dana's close is long because the acquisition added a fifth entity to a process built for four, not because her team is slow; Cadence removes the manual inter-entity reconciliation that grows with each entity. Confidence: high — rests on her own post describing a five-week close "since the acquisition".
Why this might be wrong — She describes the problem as temporary ("since the acquisition"), which suggests she may be waiting for it to settle rather than shopping. If the acquisition closed more than two quarters ago, the window has probably passed.
Opening message — "Dana — your post on the five-week close after the acquisition. The bit I keep seeing: entity count, not team size, is what sets close length. Curious whether the fifth entity added days to reconciliation specifically, or across the board?" (243 characters)
Illustrative, not a promise: this is the shape of a good answer rather than a guarantee of one, and the research in it came from text pasted into the conversation rather than from a page the model opened. The same contrast on a connection note — a checkable reference against a message that could be sent to anybody — is on the outreach resources, where Connection Request Generator writes three variants and tells you which to send.
Every LinkedIn prompt on the web is written as if the platform had no limits
Ask any assistant for a LinkedIn outreach plan and you will get one that sends fifty connection requests a day. No plan on LinkedIn permits that, and the account that tries it gets throttled — quietly, by having its own ceiling lowered. The model is not lying; it simply has no reliable knowledge of the current limits and no way to tell you that.
Applies on every plan including Recruiter Corporate. LinkedIn adjusts it per account based on acceptance rate, so a low-acceptance account gets less, never more.
Unused credits expire. They accrue to at most three times the monthly allowance, and a credit is returned if the recipient replies within 90 days.
Free accounts hit a monthly ceiling on profile searches. Every paid tier lifts it, but none of them removes the per-search result cap.
Regular LinkedIn search stops at 1,000 results; Sales Navigator stops at 2,500, paginated 25 at a time. Both are per search, not per day.
Those 4 apply on every plan, free or paid, and no tool or server in this section lifts a single one of them. They are generated here from the same source as the limits page, compiled into the prompts that plan volume, and shipped as LINKEDIN-LIMITS.md so an assistant can read them on every turn instead of guessing.
What this section does that an “AI tools for LinkedIn” roundup does not
Every row below is a structural fact about how these pages are built, checkable by opening any prompt, tool or server record — not a claim about one named competitor. It is the pattern across the category.
| This section | A typical roundup | |
|---|---|---|
| Where the numbers come from | One dated registry per subject, each read on a stated day | Unsourced, with no date on any figure |
| Platform send limits | Compiled into every planning prompt, from the pricing pages’ own data | Not mentioned — the plan assumes unlimited sending |
| What a tool does to your account | Stated per tool, derived from how it reaches LinkedIn | Not an axis at all; tools are ordered by feature count |
| Tools left out | 3 named, each with the rule it failed | Only what made the list |
| Vendors with no published price | Marked as such — 7 of 42 | A “from” figure, or a blank cell |
| Reusing your context | 15 .md files, attached once to a Project | Retype your ICP into every new chat |
| What it costs you | Free, most of it with no email, and no affiliate links | Free to read; a gated pack or a list to join |
Nothing here says “tested”
Written to a fixed structure and run through Claude and ChatGPT to confirm each one returns the output format it declares. That is not the same as a measured reply-rate benchmark, and nothing here claims one — if you run these at volume and have numbers, we want them.
Each registry carries the day it was last read and moves on its own clock. Re-reading a vendor's pricing page tells you nothing about LinkedIn's own features, so the two dates are separate and neither is allowed to restamp the other.
Every figure on this page is read from the registry that owns it. A number cannot be flattering in prose and correct in the table, because there is only one of it.
No reply-rate claims and no “increases meetings by 40%”. The tool and server directories are sorted, not ranked: we have not run 42 products against each other and neither has anyone publishing a top ten.
Use the prompts and files commercially and for clients without asking. The only thing you may not do is republish the library as your own. Full licence.
The Starter Pack — 12 files, 12 prompts, one zip
Everything most people need, assembled. It costs an email address, and the ungated link is under the form because every file in it is readable free on this site anyway.
The questions people arrive with, answered directly
Can ChatGPT or Claude read a LinkedIn profile from a link?
No. LinkedIn blocks assistant crawlers, so a model given a profile URL will either say it cannot open it or, much more often, invent the contents of the page and present them confidently. Every prompt here is written to be fed pasted text for exactly that reason, and the context files instruct the assistant to say so rather than guess. The tools and MCP servers in this section are the other answer to the same obstruction: they get in with a credential rather than with a crawler, and that credential is what the directories sort them by.
What is the best AI tool for LinkedIn?
There is no one answer, and a ranked list that gives you one is ranking on something it has not told you. The 42 tools in the directory divide first by how they reach LinkedIn: 9 do not touch it at all and cover email and the phone instead, 18 act under your name, and 12 hold your signed-in session on their own machines. Decide which of those you are willing to accept, then compare features inside that group. Every record was read from the vendor’s own pages on 20 September 2026, and 3 researched tools are listed as excluded with the rule each one failed.
Does LinkedIn allow AI tools?
Its User Agreement bars automated access by any means and carves out nothing for a well-paced tool, and its Prohibited Software policy states that accounts using third-party automation may be restricted — so the consequence lands on your account rather than on the vendor you paid. Pasting text into an assistant is not automated access of LinkedIn and sits outside that clause; a tool that signs in as you sits squarely inside it. This section describes what each route is and what it exposes, and recommends automating nothing.
What AI does LinkedIn itself include?
29 features across 6 of its products, 5 of which need no plan at all. The cheapest plan that adds any is Premium Career at $39.99 a month, read from LinkedIn on 10 September 2026. 10 of the 29 have no price that can be looked up anywhere, for four different reasons, and 4 are documented two ways by LinkedIn itself.
Is there an official LinkedIn MCP server?
No. LinkedIn and Microsoft were both checked on 20 September 2026 and neither publishes one. All 13 that work are community-built, 4 of them put no LinkedIn account behind them at all, and the one running on LinkedIn’s own OAuth cannot search people or open another member’s profile — not because its author stopped early, but because no LinkedIn permission grants that.
How do I use Claude for LinkedIn sales?
Create a Claude Project, attach the context files — ICP.md, PRODUCT.md, OUTREACH.md, LINKEDIN-LIMITS.md — and paste CLAUDE.md into the custom instructions. After that every conversation in that Project already knows who you sell to, what you may claim and what the platform permits, so you stop re-explaining your business on every request. The Sales Agent pack ships that folder ready to fill in.
What is the best prompt for LinkedIn prospecting?
The one that knows what you can actually do afterwards. A prospecting prompt that returns 400 people to contact this week is useless, because LinkedIn caps connection invitations at ~100 per week on every plan including the most expensive one. The prompts here carry those ceilings, so the list you get back is the size of a week rather than the size of a search. Start with the ICP Builder, then the Sales Navigator Search Builder, then the Lead List Qualifier — in that order, because each one consumes the previous one’s output.
Is any of this paid?
Not today. Every prompt, file and pack in this section is free and needs no account. One bundle asks for an email address — and the same zip downloads without the form, from the top of that page, because everything in it is readable on the site anyway. The tool and server directories carry no affiliate links and take no payment for placement. If a paid tier appears later it will be for the maintained, versioned system rather than for access to text that is already public.
Decide the route before you decide the tool
Nothing you paste into an assistant can get a LinkedIn account restricted. 18 of the 42 tools here act under your name and can. That decision sits upstream of every feature comparison, and this section is sorted for it.