How to connect LinkedIn to ChatGPT and Claude (2026)

You can already paste a prospect's profile into ChatGPT and ask it to write a connection request. What most people actually want in 2026 is different: an AI assistant that can read your LinkedIn, decide what to do, and then do it - send the request, follow up on replies, publish the post, pull the analytics - without you copying anything by hand. That is the difference between an assistant that drafts text and an agent that runs a workflow.

The catch is that LinkedIn is not like Gmail or Google Calendar. There is no open, personal API you can hand an OAuth token to. LinkedIn's official APIs (Marketing and Sales Navigator) are gated, partner-only, and cover almost none of the day-to-day actions an SDR or founder actually takes. So when someone asks 'how do I connect LinkedIn to ChatGPT,' the honest answer is: it depends entirely on which of three approaches you use, and they are not equal on safety, capability, or maintenance.

This guide walks through all three - copy-paste, browser scrapers, and a purpose-built MCP server - so you can pick the one that fits. We build the third, so we will be upfront about where the first two are genuinely fine and where they fall down.

Approach 1: copy and paste. This is where everyone starts, and for light work it is completely reasonable. You copy a profile, a message thread, or a job post into ChatGPT or Claude, and you ask for a draft. The assistant is excellent at the language part: rewriting a cold opener, summarizing a long thread, adapting tone. If your need is 'help me write better,' you do not need to connect anything at all - the chat window is the whole product.

Where copy-paste breaks down is volume and action. The assistant cannot see your inbox, cannot send anything, cannot check who accepted yesterday, and cannot run on a schedule. Every step is manual, and the context resets each time. It is a writing aid, not an operator. The moment you want the AI to actually take an action on LinkedIn, copy-paste is a dead end.

Approach 2: browser extensions and scrapers. The next rung is a tool that automates your logged-in LinkedIn session - a Chrome extension or a desktop app with an embedded browser that clicks around on your behalf. Some of these now bolt on an 'AI' layer. This works, and it can do real actions, but it comes with real costs. The automation fires from your own browser session, so it looks like traffic from your device and your browser has to stay open. Selectors break whenever LinkedIn ships a UI change. And crude, aggressive scraping is exactly what gets accounts restricted - LinkedIn is far better at detecting mechanical, browser-driven patterns than it was a few years ago. For an AI agent, this is also the wrong shape: the agent has no clean, stable set of actions to call, just a fragile screen it has to poke at.

Approach 3: a purpose-built MCP server. This is the approach designed for exactly this problem, and it is the one that has matured fastest in 2026. MCP, the Model Context Protocol, is an open standard for connecting AI assistants to external tools and data through a stable, described interface. Instead of the model guessing at a web page, it sees a clean list of typed actions - send a connection request, list new messages, publish a post, run a search - and calls them directly. Anthropic introduced MCP, and it is now supported across the ecosystem, including by Claude and by ChatGPT's connector and developer modes.

An MCP server for LinkedIn is a service that speaks this protocol and safely translates the agent's intent into real LinkedIn actions. That is what Crispy is: a cloud LinkedIn MCP server that exposes the complete LinkedIn surface to AI agents. Your assistant connects once, then has a described toolset it can reason about and call - no scraping, no brittle selectors, no browser to keep open.

Why MCP is the right abstraction. An agent is only as good as the tools it can call reliably. A stable, typed interface means the model can plan: 'first search for these people, then for anyone who accepted, send this follow-up, and flag replies that sound like buying signals.' Each of those is a discrete tool call with a predictable result, so the agent can chain them, handle failures, and explain what it did. A screen-scraper gives the model none of that - which is why agent-native tooling, not extensions, is where this is heading.

How to connect it to ChatGPT. In 2026, ChatGPT supports MCP through connectors and its developer/agent modes. The flow is: create your Crispy account and connect your LinkedIn account to Crispy (this happens in the cloud - there is no extension to install), generate a scoped API key, and add Crispy as an MCP connector in ChatGPT using that key. From then on, ChatGPT can call the LinkedIn tools directly in a conversation or an agent run. You describe what you want in plain language; it makes the calls.

How to connect it to Claude. Claude supports MCP natively in Claude Desktop and through connectors on claude.ai. You add Crispy's MCP endpoint and your API key as a connector, and Claude gains the same LinkedIn toolset. Because MCP is a shared standard, the exact same Crispy connection works across Claude, ChatGPT, and any other MCP-capable client - you set it up once, not once per assistant.

What the agent can actually do once connected. This is where the full-surface point matters. Crispy is not just outreach: the agent can send and personalize connection requests and messages, read and triage the inbox with urgency scoring, run people and Sales Navigator searches, publish and schedule content, analyze post performance, manage contacts and lists, and pull network analytics - all as described in the complete MCP guide. The whole point of handing LinkedIn to an agent is that it can move across all of those in one workflow, not just fire messages.

Safety is not an afterthought here - it is the reason to prefer this approach. Because Crispy runs in the cloud and enforces per-category daily limits on the server side, the agent physically cannot blow past safe activity levels, no matter how enthusiastically it plans. That is a categorically different safety model from a browser extension that executes whatever it is told inside your live session. If you are going to let an AI act on an account that matters, you want the guardrails enforced below the agent, not left to the agent's judgment. We wrote about the specifics in MCP + LinkedIn: how AI agents automate outreach.

Your data stays yours. A connected agent needs memory - who you talked to, what was said, what is scheduled - but that should not mean your LinkedIn data is locked in someone else's database forever. Crispy lets you export or delete everything at any time, so the convenience of an agent that remembers context never costs you ownership of the underlying data.

Building your own agent? Use the REST API. If you are not working inside ChatGPT or Claude but building a custom agent - in a framework, a workflow tool, or your own code - Crispy exposes the same capabilities over a REST API. Anything the MCP tools can do, your agent can do with a normal HTTP call, which means LangChain-style agents, internal automations, and no-code workflow builders all connect the same way.

A concrete example of what this unlocks: 'Every morning, check who accepted my connection requests, send each of them a short, personalized message referencing their most recent post, and put anyone who replies with a question into a list called Hot.' With copy-paste that is an hour of manual work; with a scraper it is a fragile macro; with an MCP server it is a single instruction an agent runs on its own, safely, every day.

Getting started is deliberately boring, which is the point. Create an account, connect LinkedIn in the cloud, generate a key, and add the connector to ChatGPT or Claude. There is nothing to install, nothing to keep running, and one connection works everywhere MCP does. From there, the interesting part is not the setup - it is deciding what you want your agent to actually do.

The short version: if you only need help writing, the chat window alone is enough. If you need an AI that can genuinely operate your LinkedIn - read, decide, and act, safely and on a schedule - a purpose-built MCP server is the approach designed for it, and it is the one that connects cleanly to ChatGPT, Claude, and whatever agent you build next.