- LinkedIn does not publish an MCP server. What exists falls into four families: posting servers built on LinkedIn's official API, community servers that drive your own browser session, MCP layers from LinkedIn API vendors, and MCP servers from outreach platforms.
- Pick by job: publishing posts fits the official route, personal research fits a community server, building a product fits an API vendor, and running prospecting from Claude or ChatGPT fits an outreach platform. Our guide to LinkedIn AI agents for outreach covers the agent side in depth.
- Connecting one is the easy part: a URL and a key pasted into Claude, ChatGPT, Cursor or another MCP client. What matters is what refuses the agent when it asks for too much.
- For outreach, Cadency's MCP server exposes campaigns, leads, the inbox, people search and sender health as tools, under the same per-sender ceilings and approvals as the app.
A LinkedIn MCP server lets an AI assistant act on LinkedIn. It exposes actions such as publishing a post, reading a profile, sending a connection request or reading replies as tools that Claude, ChatGPT, Cursor or any client speaking the Model Context Protocol can call. LinkedIn does not publish one, so every server you can connect today comes from a third party, and they differ less in how you connect them than in what they reach: your own posts, your own browser session, a developer API, or a full outreach engine with limits the agent cannot override.
This guide maps the options as of October 2026, shows what connecting one looks like, and lists the checks we would run before trusting one with an account that matters. It comes from the team that builds the engine underneath Cadency's MCP server, so the outreach sections are written from running the engine, not from reading changelogs.
What is a LinkedIn MCP server?
MCP, the Model Context Protocol, is the open standard Anthropic introduced in late 2024 for connecting AI assistants to external tools and data. A server declares tools with typed inputs; a client such as Claude, ChatGPT or an agent framework discovers them and lets the model call them mid-conversation. The same protocol works across vendors, which is what makes it useful: one server, every assistant.
A LinkedIn MCP server applies that to LinkedIn. Instead of a chatbot that talks about LinkedIn, you get an assistant that acts on it: the tools wrap real calls, the account doing the work is a real member, and the results come back as structured data the model can reason over. What sits behind those tools decides everything else, which is why the four families below behave so differently once the novelty wears off.
Is there an official LinkedIn MCP server?
No. LinkedIn does not publish an MCP server, and its official APIs cover a narrow slice of what people want an agent to do. Posting on a member's behalf is open to any developer app through the Share on LinkedIn product. Company pages, advertising and messaging sit behind programs LinkedIn approves case by case. Prospecting actions such as connection requests are not offered at all.
That has a consequence worth stating plainly. Every server that reads profiles, sends invitations or sends messages through a real account is automating it, which LinkedIn's User Agreement does not permit. The honest comparison between those servers is who carries that risk and how it is managed, which is the subject of our guide to whether LinkedIn automation is safe. For what the official programs do allow, see our LinkedIn API guide.
Which LinkedIn MCP server should you use?
Start from the job, not the vendor. The four families overlap less than their landing pages suggest:
| Family | How it works | Examples | Fits | You handle |
|---|---|---|---|---|
| Posting servers on the official API | OAuth into LinkedIn's API; tools publish your own content | Merge's LinkedIn connector | Publishing posts from an assistant | The connected app and its LinkedIn permissions |
| Community servers on your own session | A local browser signed in to your account, driven by the server | linkedin-mcp-server on GitHub | Personal research and light use of one account | Session upkeep, pacing, breakage when LinkedIn changes its pages, and the account at risk |
| LinkedIn API vendors | An MCP layer on a developer API that runs actions on connected accounts | Linked API, LinkupAPI, Unipile | Developers wiring LinkedIn into their own product or agent | Sequencing, reply handling and, depending on the vendor, pacing |
| Outreach platforms | An MCP layer on a finished campaign engine | Cadency, HeyReach | Teams running prospecting from Claude or ChatGPT | Choosing targets and approving what goes public |
Posting servers: the official route
If the job is publishing, start here. Hosted connectors such as Merge's LinkedIn MCP connector run the OAuth flow, store and refresh the token for you, and expose tools like create_text_share, create_article_share and create_image_share. Because they run on the API LinkedIn provides for exactly this, they carry far less account risk than the other families. The trade-off is reach: they publish posts and read basic profile or page data, and they will never send an invitation or read your inbox.
Community servers: your own session, your own risk
The most visible project on GitHub, linkedin-mcp-server, shows what this family has become. It drives a browser on your own machine, signed in to your own LinkedIn account, and exposes profiles, companies, jobs, the inbox, messages and connection requests as tools; its README lists connect_with_person and send_message next to the search tools. Getting to the first successful call takes minutes to an hour. What follows is the actual product: keeping the session alive as LinkedIn changes its pages, deciding how fast the agent may act, noticing when a tool breaks silently, and accepting that the seat at risk is yours.
None of that is a criticism. For one account used gently, a community server can be all you need, and it is the best way to learn how LinkedIn behaves. The line we draw is commercial: once the outreach feeds a pipeline someone is paid to fill, pacing, monitoring and recovery stop being a hobby and become an unpaid second job.
LinkedIn API vendors: a chat interface on raw actions
Developer APIs that run LinkedIn actions on connected accounts now ship MCP servers too. Linked API offers one for Claude, ChatGPT, Cursor and Codex on top of its per-seat API. LinkupAPI markets its MCP for LinkedIn, email and WhatsApp outreach, where the agent searches prospects, sends messages and reads replies. Unipile takes a different angle: its MCP server serves coding agents that build LinkedIn features into a product, rather than an assistant running outreach from chat.
The common thread is that these are raw actions with a chat interface. A tool sends the message you ask for; the sequence around it, the reply handling and, depending on the vendor, the pacing between calls stay in your prompt or your code. That is the right shape when you are building something yourself, and a heavy one when all you wanted was to run a campaign.
Outreach platforms: a finished engine behind the tools
The fourth family puts the MCP server on top of a campaign engine, so the agent drives campaigns, leads and the inbox rather than individual sends. HeyReach markets its MCP for launching campaigns, drafting messages, assessing performance and acting on warm leads from chat.
Cadency's MCP server exposes the whole product as tools. The assistant can search people and companies and enrich them, create campaigns with their steps, delays and senders, enrol, pause or remove leads, read conversations filtered by interest and answer them as the right sender, check each account's health and remaining ceiling, and pull reply and acceptance rates by step, sender and campaign. Cadency began as our internal test bed on the Swarmhit Smart API, and its MCP server has no special powers: a call from an assistant is a call from the API key's role, under the same ceilings, roles and approvals as the app.
How do you connect Claude or ChatGPT to a LinkedIn MCP server?
For a hosted server, the setup is the same everywhere: create a key on the provider's side, then give your client the server URL and the key. In Claude, that means adding a custom connector. ChatGPT connects to remote MCP servers as apps in developer mode, on the web, and which plans may use them for write actions has changed more than once, so check your workspace settings before you plan around it. Cursor, Claude Code and agent frameworks read the same URL from their MCP configuration, and n8n can call an MCP server inside a workflow.
On Cadency, the four steps are: create an API key, one per workspace, scoped by role and revocable at any time; add the server URL and the key to your client; ask in plain language; and review what it did, because every call is logged like an action in the app. Self-hosted servers add a local install and a sign-in step.
What a first outreach session looks like
A first session tends to follow the same arc. You ask for heads of sales at German SaaS companies; the assistant searches and returns a table you can argue with in plain language, tightening the filters twice without touching a screen. You describe the sequence you want, a short intro, a three-day wait and a follow-up, and name the three senders who should run it. The assistant creates the campaign, enrols the leads and launches it once you confirm. A week later, one prompt returns who replied with interest, and the assistant offers to draft answers as the right sender.
The part worth watching is not the speed. It is what happens when you ask for more than the accounts can take: the answer should be a refusal from the server, not an apology from the model after the damage. The model proposes, the infrastructure disposes.
Five workflows that earn their keep
Once the novelty wears off, these are the patterns that keep an MCP connection open on a normal week:
- The morning triage. One prompt: what replied overnight, and what needs me? The assistant separates interest from polite declines and drafts answers for the threads worth a human. A morning of inbox triage shrinks to a few minutes.
- The agency review. One connection per client workspace, one prompt across them: campaign stats, sender health, anything paused or failing. A Monday of tab-switching becomes a briefing.
- The pre-call brief. Before a demo, ask for everything known about the prospect: past conversations from the inbox, profile context, which campaign touched them and how they reacted.
- The capacity check. Before asking for more volume, ask which senders have room left today and which are close to their ceiling. The answer comes from the server, not from a guess.
- The inbound loop. An n8n workflow catches a form fill or a trial signup, checks whether the person is already in a campaign and adds them to the right one if not. Our n8n and LinkedIn guide shows the API version of the same loop.
What should a LinkedIn MCP server refuse to do?
Language models are enthusiastic. Given a goal and a send tool, one will cheerfully burn a seat in an afternoon. So the question to ask of any LinkedIn MCP server is not what the agent can do, it is what the infrastructure refuses to let it do. On Cadency the refusals are structural: per-sender ceilings on every action, campaign steps paced by the engine, a dedicated proxy for each sender, and checkpoint handling, with no setting that lifts a ceiling past the platform maximum. An assistant cannot exceed a sender's ceiling, skip an approval, or read a workspace its key does not belong to.
Two habits complete the picture. Treat MCP keys like production credentials: one key per workspace, so an agent never wanders across clients, and revoke a key the moment a laptop or a contractor leaves. And keep public content human: in Cadency, AI comment steps queue for a person to approve, so route anything the assistant writes for public view through that queue, which is exactly the distance a second pair of eyes deserves. If an account does get restricted, our guide to a restricted LinkedIn account covers the recovery.
How to evaluate a LinkedIn MCP server
If you are comparing options this week, here is the test we would run, in order.
- Coverage. List the tools and check they span the whole job: search, audience, campaigns, inbox, reporting. A server that cannot read replies turns your agent into a spammer with amnesia.
- Enforcement. Ask where the limits live. The only acceptable answer is on the server. If the documentation suggests configuring the agent to behave, the safety layer is your prompt, and prompts do not survive a confused model.
- Approvals and scoping. Anything AI-written that becomes public should pass a human queue, and multi-client setups need per-workspace keys you can revoke one by one.
- Parity. A server that mirrors a documented API gives you a path to code later; one that exposes a different, undocumented surface gives you a science project.
- The mundane checks. Does it work with the clients your team already uses, is the pricing written down, and is every call logged where you can read it?
Should you build your own LinkedIn MCP server?
If you are building an agent product, the MCP server you ship to your users is yours to design, and the real question is what sits under it. Wrapping raw actions gives your agent the same problem a raw API gives a developer: it can send, but nothing sequences, paces or refuses. The Swarmhit Smart API is built for that layer. One call creates a campaign with its whole sequence, then you add leads and activate it; per-sender limits are enforced server-side on every action; campaigns are paced by the engine; and replies arrive as webhooks your agent can act on. Its direct actions run under the same per-sender caps, but the spacing between those calls is yours to set.
Whether your agent should talk MCP or your code should talk REST is a separate decision, and our MCP vs API guide walks through it.
FAQ
Is there an official LinkedIn MCP server?
No. LinkedIn does not publish an MCP server, and beyond sign-in its official APIs open only posting to every developer, with company pages, ads and messaging behind approval programs. Every LinkedIn MCP server you can connect comes from a third party: hosted posting connectors on the official API, community projects that drive your own browser session, MCP layers from LinkedIn API vendors, and MCP servers from outreach platforms such as Cadency or HeyReach. Pick the family by the job you need done.
What is the best LinkedIn MCP server?
It depends on the job. To publish posts from an assistant, a posting connector on LinkedIn's official API is the safe choice. To research with your own account, a community server such as linkedin-mcp-server on GitHub is free and capable, at your account's risk. To build LinkedIn into your own product, an API vendor's MCP layer fits. To run prospecting campaigns from Claude or ChatGPT, an outreach platform such as Cadency keeps sequencing, campaign pacing and per-sender ceilings on the server.
Does ChatGPT support LinkedIn MCP servers?
Yes. ChatGPT connects to remote MCP servers as apps in developer mode, on the web, and the server's tools then appear in the conversation. Which plans may run write actions, such as sending a message, has changed more than once, so check your workspace settings before relying on it. Claude supports remote MCP servers as custom connectors, and Cursor, Claude Code and n8n can use the same server URL.
Do I need to code to use a LinkedIn MCP server?
No, not for a hosted server. Setup is creating a key on the provider's side and pasting the server URL and the key into your client. Self-hosted community servers need a local install and a sign-in step, which is light work for a developer. Code only enters the picture when you want the same actions inside your own product, which is what a REST API and webhooks are for.
Can a LinkedIn MCP server send connection requests?
Some can. Posting connectors on LinkedIn's official API cannot, because that API offers no invitation endpoint. Community servers such as linkedin-mcp-server send them through your own browser session with a connect_with_person tool. API vendors expose them as single actions on connected accounts, and outreach platforms such as Cadency send them as campaign steps, capped per sender by the server. All of them automate a real account, which LinkedIn's User Agreement does not permit.
How much does a LinkedIn MCP server cost?
Community servers are free to run, but the account at risk is your own. API vendors price their MCP access like their API, usually per connected account or per seat. Outreach platforms bundle it with the product: Cadency includes its MCP server and its API in every plan, from $89 a month billed monthly with 2 LinkedIn senders on Essential, $199 with 5 on Small Team or $699 with 20 on Scale; enrichment runs on credits, charged per verified email found. Posting connectors follow their provider's own pricing.
