Best LinkedIn MCP Servers in 2026 (Sorted by What They Let an Agent Actually Do)

In August 2026 we asked GPT-4o, with web search turned on, which LinkedIn MCP server to use for running outreach with an AI agent. It named eight products. Every single source it cited was a vendor’s own website. Not one review site, not one forum thread, not one independent test.

That tells you something useful about the state of the LinkedIn MCP category: it is young enough that the vendors are still writing the map. So here is a map with the measurements shown. This post sorts twelve LinkedIn MCP servers by the one property that actually decides whether a given one fits you, which is not the tool count and not the price. It is what sits between your agent and your LinkedIn account.

How big is this actually, and who currently ranks

Two numbers worth having before you read anyone’s recommendation.

Demand is real and rising.Google Ads data pulled through DataForSEO puts US searches for “linkedin mcp” at a 1,000 per month average over the trailing year, with 2,400 in June 2026, the most recent month reported. “Linkedin mcp server” adds another 480 per month. Google Ads publishes volumes on a one to two month delay, so treat June as the freshest reading rather than today’s.

Nobody owns the answer yet.We pulled the live US desktop results for “linkedin mcp” on 10 August 2026. Position one is a LinkedIn post announcing an open-source release. Position two is a GitHub repository. Position three is a Reddit thread. The rest of page one is directory listings: mcpservers.org, Apify, Zapier, and the Docker Hub MCP catalog. Google is also rendering an AI Overview on the query. In a mature category the vendors hold those slots. Here, the community does.

The question that sorts every server on this list

When an agent decides to contact 200 people, what happens next? There are exactly three answers in this market, and they split the products cleanly.

1. Nothing, because the server cannot contact anyone.It is built on LinkedIn’s official API, which does not expose outreach.
2. 200 requests go out, roughly now. The agent holds the send button.
3. 200 people enter a queue and get contacted over the coming weeks. The agent wrote a plan; something else decides the pace.

Tool counts vary by a factor of 80 across this list, and that variation mostly tracks which group a server is in rather than how good it is. Read the groups, then pick.

Group one: official-API servers (compliant, and cannot do outreach)

These connect through LinkedIn’s documented REST API with OAuth. They are the only options here that are unambiguously within LinkedIn’s rules.

• Zapier MCP. Two LinkedIn write actions, verified on their page: create a company page update, and share a status update. That is the whole surface.
• Composio. A managed-OAuth layer advertising 22 LinkedIn tools, priced per tool call. The broadest official-API option.
• souravdasbiswas/linkedin-mcp-server. Open source under MIT, 14 tools, OAuth against the official API. Posts and events.
• Taplio. 14 tools focused on content, gated behind a Taplio subscription.

The ceiling is structural, not a product decision. LinkedIn’s open permissions cover profile data, email, and posting, commenting, and liking as the signed-in member. The Messages API is restricted to approved partners and, even then, only allows member to member messages to first-degree connections or replies inside an existing thread. Developers publicly report partner approval taking one to four months, with use cases that touch messaging or bulk profile data among the common rejection reasons.

Choose this group ifyour agent’s job is publishing and reporting and you want zero account risk. Skip it ifthe word “outreach” is anywhere in your requirements.

Group two: direct-drive servers (the agent gets the send button)

These operate LinkedIn the way you do, through a browser session or unofficial endpoints. The agent calls a tool named something like send_connection_request and it happens at that moment.

• Crispy.The most capable server in this category by a wide margin. Its pricing page advertises “All 175 tools, no feature gates” at $49 per seat per month, where a seat is one connected LinkedIn account. Search, inbox, campaigns, content, analytics, and Sales Navigator, in Claude, Cursor, Windsurf, n8n, or over REST. If you want the widest possible agent surface, this is the honest answer.
• stickerdaniel/linkedin-mcp-server. The one ranking second on Google. Apache 2.0, free, 20 tools including people search, inbox reading, and send_message. It drives a real Chromium browser through Patchright with your persistent session cookie. Worth noting that several published comparisons list this project at 15 tools; we counted 20 in the repository on 10 August 2026.
• Linkydy. Executes through a Chrome extension inside your own browser session, with no artificial daily cap and a get_safety_status tool you are expected to consult yourself.
• Apify and Bright Data. Scraping actors for profile, company, and job data, billed per result. Research inputs rather than outreach.

The trade is straightforward and worth stating plainly: this group gives a language model the pacing decision. Models are not calibrated on LinkedIn’s invisible thresholds, and an agent that concludes 200 invitations is a reasonable Tuesday will attempt 200 invitations. The maintainers of the open-source option say so directly in their own README, warning that LinkedIn’s User Agreement prohibits automated access, that accounts using automated tools can be restricted or banned, and that there is no guarantee of account safety. That is a more honest disclosure than most commercial vendors in this category offer.

Choose this group if you want maximum capability, you understand the risk, and you plan to supervise. Skip it if the account in question is the one your pipeline depends on.

Group three: queued servers (the agent plans, infrastructure sends)

Here the MCP tools write campaign state and nothing else. A separate system performs the LinkedIn activity later, applying daily caps, working hours, warm-up, and cooldowns.

• HeyReach. Built for agencies running many sender accounts, with rotation across them as the headline capability. Roughly 25 tools, and the MCP itself costs nothing on top of a subscription. If you manage LinkedIn for a dozen clients, this is the one designed for your problem.
• FirstTouch. Logs every agent action into HubSpot, with human-in-the-loop approval, pacing controls, auto-withdrawal of stale requests, and SOC 2 Type II. The pick when attribution and compliance paperwork matter more than tool breadth.
• Salesforge, lemlist, Clay, Saleshandy. Multichannel platforms where LinkedIn is one channel among email and calls. Clay is upstream of all of it: research and enrichment rather than sending.
• Pro Reach (disclosure: our product). 8 tools, hosted, no extra cost on any plan. The narrowest surface on this page, deliberately: the agent has no tool that sends anything, no tool that accepts LinkedIn credentials, and launching a campaign requires a second confirmed call after a preflight report.

The cost of this group is real and you should weigh it. You lose keyword people search from chat, you lose instant sending, and day one looks quiet because the pacing engine is doing its job. What you get is a bounded failure mode: a bad agent decision produces a badly targeted campaign you can pause, rather than a restricted account.

If that division of labour is the one you want, the full tool reference and setup snippets take about five minutes to work through.

Side by side

Figures are as published by each vendor or repository on 10 August 2026. Tool counts move fast in this category; check before you commit.

Picking one without reading twelve pricing pages

Answer these in order and you will land on a group, which is 90 percent of the decision.

Does your agent need to contact people at all? If no, take an official-API server and stop reading. You get compliance for free and lose nothing you wanted.

Is the LinkedIn account replaceable? A throwaway research profile changes the maths completely. If losing it would cost you nothing, direct drive gives you far more capability for the same money.

How many accounts are you running? One or two is a solo problem. Twenty client accounts is a rotation problem, and that is a specialist category where the agency-focused tools are simply better built than the rest of this list.

Who should absorb a bad call, the model or the vendor? This is the real question underneath all of it, and neither answer is wrong. Some people want an agent that does exactly what it is told at the speed it decides. Others want the machine to refuse.

The terms-of-service part most comparisons leave out

LinkedIn’s User Agreement prohibits third-party software that automates activity on the platform. Connecting through MCP rather than a Chrome extension changes the plumbing, not the rule. Any page that lists these servers without saying so is selling you something.

What follows from that is a posture, not a verdict. Group one carries no account risk because it does nothing risky. Groups two and three both carry some, and pacing reduces it without eliminating it: a vendor can spread your activity across weeks and stay inside published limits, but no vendor can promise LinkedIn will never act. The sensible sequence, and the one the more careful projects in this space recommend themselves, is to start with reading, research, and drafting, add messaging next, add connection requests last, and keep volumes conservative for the first month on any account you care about.

The category will look different in six months. Right now the open-source projects are outranking every commercial vendor on the main keyword, which does not usually last. What will not change is the question at the top of this post, because it is the only one that survives a product roadmap.

Frequently asked questions

Related reading

• LinkedIn MCP server reference Every tool, the auth model, and copy-paste config for Claude Code, Cursor, and Claude Desktop.
• Turn your AI agent into an AI SDR A real seven-turn agent session, from a vague prompt to a preflight report.
• How to write a LinkedIn post that gets engagement The first-line discipline that decides whether a connection note gets accepted.