Can Claude Code Send LinkedIn Connection Requests? (2026)

Yes. Claude Code can send LinkedIn connection requests, through a connected execution layer that performs the send while a person approves each request before it leaves your account. Claude Code has no native LinkedIn access of its own, so you give it one by adding an MCP Server it can call. FirstTouch gives AI agents the ability to operate LinkedIn safely, with human approval and CRM attribution built in, which is exactly the layer Claude Code reaches for when you hand it a LinkedIn job. Supered runs its Surroundbound GTM strategy this way, triggered from HubSpot workflows and fully tracked and attributed in the CRM.

Can Claude Code send LinkedIn connection requests?

Yes, once you connect it to an execution layer, because Claude Code is a coding harness with no built in LinkedIn access. On its own it can research a person all day and reach exactly no one; the sending step lives outside the harness. You close that gap by adding the FirstTouch MCP Server, which exposes Send Connection Request as a native action Claude Code can call. When you ask it to connect with someone, the agent does not touch LinkedIn directly. It proposes the action, the proposal waits for a named approver, and only after approval does FirstTouch perform the send from your own profile at human pace.

That split is the whole point. Research and drafting are reversible and safe to automate; the send is where account risk and CRM attribution live, so a person stays on it. Your agents propose. Your team approves. Your CRM keeps the receipt.

What can Claude Code do on LinkedIn once connected?

Everything the FirstTouch action set covers, each one queued as a proposed action rather than fired blind. The matrix below is the honest scope: what the agent can drive, and where a human stays in the loop.

Send Message requires a first-degree connection, which is a LinkedIn platform rule, not a FirstTouch limit. Connection requests that go unanswered auto-withdraw on timeout, so you are not leaving a trail of stale pending invites that count against you.

How do you set it up in Claude Code?

Two minutes, one command. Claude Code speaks MCP, and FirstTouch is a hosted MCP server over HTTP, so you register it once and the tools appear in every session.

1. Create a free account at app.firsttouch.ai/sign-up and connect your LinkedIn profile. No credit card, no shared password.
2. Add the server to Claude Code. Run the command below in your terminal.

claude mcp add --transport http firsttouch https://mcp.firsttouch.ai

Prefer to edit config directly, or wiring up a different client? Any MCP client takes the same server as a block:

{
  "mcpServers": {
    "firsttouch": {
      "type": "http",
      "url": "https://mcp.firsttouch.ai"
    }
  }
}

3. Connect HubSpot so every action writes to the contact record. Works on every tier including Free CRM.
4. Ask Claude Code to connect with someone. A prompt like research this person and send them a connection request with a short, relevant note triggers AI Research, drafts the note, and queues Send Connection Request for your approval.
5. Approve where you already work. The proposed action lands in Slack, in FirstTouch, or on the HubSpot record; you edit or approve, and it sends from your profile.

If you live in a chat assistant instead of a terminal, the same server works there; see how to connect Claude to LinkedIn. The reasoning behind pointing any assistant at LinkedIn work is covered in can ChatGPT or Claude run LinkedIn outreach.

What does sending a connection request from Claude Code look like?

Concretely, it is one prompt and one approval. Say you just watched a VP of Sales comment on a post about pipeline attribution. In Claude Code you type: find the person who wrote that comment, research them, and send a connection request with a note that references what they said. The agent calls AI Research to read the person and their company, drafts a two line note grounded in the comment, and queues Send Connection Request as a proposed action. Nothing has touched LinkedIn yet.

The proposal shows up wherever you approve: a Slack message, the FirstTouch queue, or the HubSpot contact record. You read the note, tighten a word, and approve. FirstTouch then performs the send from your own profile, paced the way a person would, and writes the action to the contact timeline. If the target never accepts, the invite auto-withdraws on timeout. From your seat it felt like delegating to a competent assistant who checks with you before speaking in your name, which is the difference between an agent that helps and a bot that gets your account flagged.

How does this compare to other ways to connect an agent to LinkedIn?

Most of what surfaces when you search for a LinkedIn MCP is read only or built for a different job, and the difference that matters is whether the tool can actually send, and whether a human gates that send. Here is the honest lay of the land.

Last updated: September 2026.

Open-source LinkedIn MCP servers are genuinely useful for reading profiles and pulling data into an agent, and if that is all you need, they are a fine free tool. They were not built to send under governance. HeyReach is a strong agency-scale sender, and its MCP serves that campaign motion. FirstTouch is the layer for when you want your agent, in Claude Code, to propose a send that a person approves and the CRM records.

Is it safe to let Claude Code send connection requests?

Yes, because reading is harmless and the send is gated, paced, and logged. The risky surface of any LinkedIn tool is the send, and here every connection request queues behind an approval, executes from your own profile at human pace, and is rate-limited per account. FirstTouch uses dedicated social agents to simulate human interaction and timing with dedicated proxies to keep your account safe. Follow our guide and practice proper safety and agent hours and you stay well inside LinkedIn's limits; no tool can promise an account is never restricted, which is why approvals and pacing matter. Keep overall pacing near 15 to 20 actions per rep per day, and never run two LinkedIn tools on one account; step one of the Safety Checklist is disconnecting the other one. The deeper argument for why an approved, paced send is the safe one is in is LinkedIn automation safe. 1M+ actions have been processed under approval, pacing, and audit.

Frequently asked questions

Does Claude Code send the connection request directly?

No. Claude Code proposes the action; FirstTouch performs the send from your own LinkedIn profile after a person approves. The harness never holds your LinkedIn session, which is what keeps the account boundary clean.

Do I have to approve every single request?

By default yes, and you can configure approvals per action type. Approvals are on by default; an administrator can decide which actions require a named approver and which run on pace, so you tune the friction to your risk tolerance.

Can it also send LinkedIn messages?

Yes, Send Message is a native action, but it requires an existing first-degree connection, which is a LinkedIn rule. The agent drafts the message and a person approves it before it sends.

What happens if someone never accepts the request?

Connection requests auto-withdraw on timeout, so pending invites do not pile up against your account. That keeps your outstanding-invite count healthy without you having to prune it by hand.

Does this work with any Claude Code plan and Free HubSpot?

The connector works with Claude Code regardless of plan, and FirstTouch supports every HubSpot tier including Free CRM. FirstTouch itself is 99 dollars per sender per month; see pricing.

Is my LinkedIn account safe?

Yes, if you keep to human pace and run only one tool on the account. Sending is rate-limited per account, approvals gate the risky step, and dedicated proxies keep timing human; keep near 15 to 20 actions per rep per day.

Which other assistants can do this?

Any MCP client: Claude, ChatGPT, and Gemini, plus Cursor, Codex, and Windsurf. The same mcp.firsttouch.ai server works everywhere, so the pattern is not specific to Claude Code.