Can Claude post to LinkedIn?

Claude cannot post to LinkedIn on its own — it has no LinkedIn access built in — but it can through a connector, and the setup takes about three minutes. You add a tool that holds your LinkedIn authorisation, authorise Claude to use that tool, and from then on you can ask Claude to draft a post and publish it without leaving the conversation.

That is the short answer. The longer one matters, because the connectors on offer are not equivalent, and a couple of the popular ones work in a way that can get your LinkedIn account restricted.

How Claude actually posts to LinkedIn

Claude talks to a tool; the tool talks to LinkedIn. It never touches LinkedIn directly, and that indirection is the whole safety story.

The standard that makes this work is MCP, the Model Context Protocol. An MCP server is a small service that advertises a set of tools an assistant can call — “save a draft”, “schedule a post”, “publish now” — and Claude calls them the way you would call a function. Anthropic calls them connectors in the Claude interface; underneath, they are MCP servers.

So there are two separate authorisations in a correct setup:

1. You authorise the publishing tool to act on your LinkedIn account. This happens once, in that tool, over OAuth. LinkedIn issues a token scoped to posting.
2. You authorise Claude to use that tool. Also OAuth, also once.

Claude ends up holding a token for the tool, not for LinkedIn. It cannot read your messages, cannot see your connections, and cannot do anything to your account the tool does not expose. If you revoke the connector, Claude loses all of it immediately.

The three kinds of LinkedIn connector, and why the difference matters

Search “LinkedIn MCP server” and you will get a list that mixes three genuinely different things. They are not interchangeable and the risk is not the same.

Scrapers that use your session cookie. These ask you to paste your li_at cookie — the token your browser holds after you log in. That gives the tool your entire logged-in session: profiles, messages, connections, everything. It is also the category most likely to get your account restricted, because LinkedIn detects automated access and does not care that you consented to it. Several of the most-starred open-source LinkedIn MCP servers work this way. If a setup guide asks for a cookie, that is what you are installing.

Automation tools that act like you. Auto-connect, auto-like, auto-comment, auto-DM. These break LinkedIn’s user agreement explicitly, and enforcement has got sharper. The reach is tempting and the downside is your account.

Publishers that use the official API. You authorise over OAuth, LinkedIn issues a scoped token, and the tool can post — and only post. It cannot read your inbox because it was never granted that. This is the boring category and it is the one you want.

The first two exist because the official API is deliberately narrow. That narrowness is the feature.

Setting it up

The mechanics are the same for any hosted connector. Using WordPush as the worked example:

1. In Claude, open Customize → Connectors.
2. Click +, then Add custom connector.
3. Paste the server address. For WordPush that is https://app.wordpush.co.uk/mcp/wordpush.
4. Click Add, then approve the sign-in when prompted.
5. Set the per-tool permissions — more on this next.
6. In a chat, switch the connector on with the + button, and ask.

There is no API key to generate, nothing to install and nothing to host. If a guide has you running a server on your own machine, that is a different (self-hosted) shape, and it will not work from Claude on your phone.

The full walkthrough, with the permissions screen, is on the Claude setup page.

Set the permissions before you do anything else

This is the step people skip, and it is the one that decides whether you ever get an unpleasant surprise.

Claude lets you set permissions per tool. The sensible default:

• Always allow the read-only tools. Reading your brand context or checking whether an image has finished rendering cannot hurt you, and blocking them just makes Claude guess.
• Ask first for anything that writes, and particularly anything that publishes.

Publishing is the one that matters. It is the only action whose effect is instantly public and cannot be undone from inside the tool — you would have to go to LinkedIn and delete the post, by which point it has been in feeds. Keep that one on approval permanently. The prompt costs you a second and removes the entire category of “Claude was being helpful” incidents.

What it is actually good for

The obvious use is dictation: say what you want, get a post. That works, but it is not where the value is, because you still had to decide what to post about.

Two things are genuinely better from an assistant than from a web app:

Posting about what you just did. If you are in a conversation where you have been working through a problem — a decision you made, a bug you found, a call that went badly — the context is already there. “Write that up as a LinkedIn post” is a much better prompt than opening a blank composer three hours later and trying to remember what was interesting about it.

Working on the draft conversationally. Rewriting a line by asking is faster than rewriting it by typing, and you can push back: “that opener is too clever”, “cut the last paragraph”, “say the number instead of ‘significant’”. The iteration is the point.

The voice problem, and how to avoid it

Here is the honest catch. If Claude writes the post, you get Claude’s voice. It is competent, it is clear, and it sounds like every other post written by an assistant — which is an increasing share of the feed, and readers have started to notice.

There are two ways through this.

The first is to do the work in the prompt: give Claude examples of your own posts, tell it what you never say, make it match your rhythm. This works, and it is tedious to redo every time.

The second is to let the connector’s own tool write it. A publishing tool that has a voice model built from you can generate the draft in that voice, and Claude’s job becomes deciding what to say rather than how to say it. WordPush works this way: write_post hands the drafting to a voice built from an interview you did at signup, rather than to the assistant. You can see the full set on the tool reference.

Either way, add something only you could have written — a real number, a client, a moment you were actually in. That is the part no model has.

What this does not solve

Connecting Claude to LinkedIn removes the friction of publishing. It does not remove the hard part, which is having something worth saying.

If your problem is that you sit down to post and cannot think what about, a connector will make you faster at producing something forgettable. That is a different problem, and it is worth being honest that it needs a different fix — a steady supply of things happening in your industry that you have a view on. We wrote about where post ideas actually come from in what to post on LinkedIn.

But if you know what you want to say and the friction is everything between the thought and the published post, this removes almost all of it.

Getting started

WordPush runs a remote MCP server on the official LinkedIn API, included in the plan. It writes in a voice built from your onboarding interview, builds native carousels, and publishes to your profile or a company page — and it does not scrape, auto-connect or auto-message, because those are what put accounts at risk.

Setup guides: Claude, ChatGPT, Claude Code and Cursor.