How to automate LinkedIn using Claude: a practical outreach workflow

You can use Claude to take a large part of the manual work out of LinkedIn outreach. Connect it to a workspace, let it inspect prospects and lists, ask it to prepare messages and a campaign plan, then review each change before it happens. What you should not assume is that a Claude connection gives you permission to scrape LinkedIn or send invitations through an automated browser. LinkedIn's published policy says third-party software and extensions that automate activity on its website are not allowed.

Here is the short version of the workflow: choose the right Claude client, connect a trusted outreach tool, verify what the connector can actually read, build a sourced list, draft a flow, and approve a specific workspace change. Decide separately whether any LinkedIn account action is permitted. That separation makes the work faster without turning a list of prepared records into a claim that people were contacted.

First, decide which kind of LinkedIn work you mean

“Automate LinkedIn using Claude” gets used for at least three jobs. Claude can help write posts, it can help prepare sales outreach, or it can be given a browser and told to click around LinkedIn. Those are not interchangeable. A post draft is content; a prospect list is workspace data; a sent invitation is an action against somebody else's account and your LinkedIn account.

This guide is about B2B prospecting and outreach preparation. Claude does the reading, sorting and drafting with a connected CRM. You keep the approval decisions. If your goal is simply to publish better LinkedIn posts, start with a writing and scheduling process instead. You do not need to give an AI agent control of a logged-in LinkedIn session just to make a content calendar.

For outreach, the useful handoff is a reviewed list: who belongs, why they belong, what evidence you have, and what should happen next. A tool that produces 200 names but loses the reason for each name has saved little work. The checking merely moves to later, when errors are more expensive.

Choose Claude.ai, Cowork or Claude Code

Claude has more than one way to use a connected tool. In Anthropic's current connector guide, custom remote MCP connectors work in Claude, Cowork and Claude Desktop. An individual account can add one from Customize > Connectors; a Team or Enterprise organization's owner adds it for the organization first. A Free account is limited to one custom connector. The connector is then enabled for the conversations where you want to use it.

If you work mainly in chat or Cowork, start with the Claude custom connector and its sign-in flow. If your prospecting process sits beside files, CSVs, internal notes or code, Claude Code may suit you better. Claude Code's MCP documentation supports a remote HTTP tool server and configuration you can inspect from the CLI. These clients can use related integrations, but do not assume a connector in one session has already been authenticated, enabled and granted the same tools in another.

The connector is a bridge to the outreach workspace. It does not connect Claude directly to all of LinkedIn's private functions. That matters when a guide promises “Claude does everything” without showing which service actually holds the lists and campaigns.

Connect Claude to OutreachGenie in the web app

OutreachGenie's live Claude guide describes the hosted MCP endpoint and its workspace tools. For Claude.ai or Cowork, use the custom connector path and OAuth sign-in when offered:

1. Open Customize > Connectors in Claude and add a custom remote connector. Name it OutreachGenie and enter https://api.outreachgenie.ai/mcp. If your Claude account is managed by a Team or Enterprise organization, its owner may need to add the connector first.
2. Connect to OutreachGenie, sign in, choose the intended workspace and read the requested scopes. If the requested access is broader than the job requires, stop and change the setup before approving it. A connector can read or change data according to what you authorize; “connect” is not a harmless view-only button.
3. Enable OutreachGenie for the particular Claude conversation. Anthropic's help page places that toggle under the + menu beside the conversation input.
4. Ask for a read-only inspection before requesting any import or campaign work.

Use a prompt that leaves little room for guessing:

Confirm which OutreachGenie workspace you can access. List the available tools and existing prospect lists. Do not create, edit, import, message or launch anything. If a list or tool is unavailable, say so rather than inventing a result.

Read Claude's answer against the workspace you expected. A sign-in success proves that authentication happened; it does not prove the assistant read the right tenant, found the right list or can perform a particular action. If the tool list is empty, check that the connector is enabled in this conversation and that you approved the intended workspace and scopes.

Connect Claude Code without putting a live key in a command

OutreachGenie's Claude Code instructions also document a workspace API-key path while OAuth connector use is being rolled out. The first-party guide shows an HTTP MCP server at https://api.outreachgenie.ai/mcp and says to create the key under OutreachGenie Settings > API keys. Keep that key out of prompts, source control, screenshots and terminal commands that will be retained in shell history.

Claude Code's own MCP guide supports ${VAR} expansion in a trusted project's .mcp.json, including the authentication header. A configuration example is:

{
  "mcpServers": {
    "linkedin-crm": {
      "type": "http",
      "url": "https://api.outreachgenie.ai/mcp",
      "headers": {
        "Authorization": "Bearer ${LINKEDIN_CRM_API_KEY}"
      }
    }
  }
}

Make LINKEDIN_CRM_API_KEY available to the Claude Code process through a private environment or secret manager. The JSON contains a variable reference, not the value. A project-scoped .mcp.json is only suitable in a repository you trust: team members who run it will connect to the named server with their own environment. Do not commit a literal token. If you use a private user-scoped configuration instead, apply the same secret rule. Anthropic's expansion notes describe what happens if a variable is missing, so check the environment before troubleshooting the tool itself.

Restart Claude Code after changing the environment. Run claude mcp list, claude mcp get linkedin-crm, or /mcp in the Claude Code session. These tell you whether the server configuration is present and whether authentication needs attention. Follow with the same read-only workspace prompt used above. That tool call is the real preflight. A server in a list is not evidence that a prospect import succeeded.

If your Claude Code account already exposes the OutreachGenie connector from Claude.ai, use the client-supported connector path rather than setting up a second key by reflex. Check /mcp and the workspace scope in the actual Claude Code session. The exact availability can depend on your Claude authentication mode and plan; Anthropic's MCP guide is the current reference.

Give Claude a list rule it can check

The quality of the first list matters more than the speed of the first send. Define an inclusion rule and an exclusion rule before asking Claude to import anyone. For example: “Current heads of revenue at UK B2B software companies with 20 to 200 employees, where we have a verified reason to discuss outbound process; exclude existing customers, prior declines and people with no current-role confirmation.” The exact market is yours, but the rule must be checkable record by record.

Ask Claude to review a small batch you already have permission to use, such as a prepared CSV or your workspace's existing prospects:

Review the current list against this audience rule. For each record, show its source, current role, company, duplicate status, exclusion status and the one reason it may fit. Mark anything you could not verify as "needs research." Do not import or contact anyone yet.

That prompt forces a useful failure mode. If Claude lacks a source, it should not turn an uncertain title into a confident personalized message. If it sees two records with the same LinkedIn profile URL, it should not create two new contacts. OutreachGenie's prospecting workflow treats collection and qualification as separate steps.

After you inspect the shortlist, approve only the proposed workspace change: which records will be added or matched, which list will receive them, and which were rejected. A note stored in a CRM is not a connection request. OutreachGenie's product page says saving a prospect does not start a campaign.

If the change starts from a CSV, preview the field mapping before import. Check that profile URL, current role, company and source evidence land in the right fields, and reject rows that cannot support the audience rule. A clean-looking list with shifted columns can be harder to repair than a small batch left for manual review.

Draft the outreach, then review the operating state

Now Claude has something real to write from: a short, qualified list with source notes. Have it draft a first message using only the verified details you supplied. Avoid “saw your exciting work” if neither you nor Claude can point to that work. One honest observation and a clear reason to talk are usually better than five lines of synthetic familiarity.

Ask for a plan rather than a launch:

Using the reviewed list, draft a LinkedIn outreach flow. Show the aim of each step, the facts used in each message, the wait or prerequisite before it runs, and the condition that stops further follow-up. Keep the campaign inactive. Show five prospect-message pairs for review, and ask before creating or changing workspace state.

Inspect what Claude returns. Does the invitation assume a connection already exists? Does a follow-up need a reply check? Are people who declined or already replied excluded? Is every claim in the copy supported by the record? If the connected tools do not expose a reply or failure state, the plan should say that you will handle that decision manually. It should not pretend to monitor a state it cannot see.

This is also where Claude's broad “do it for me” strength needs a boundary. Approve one named list and one named flow, not a general instruction to run outbound. Review the campaign preview in OutreachGenie before changing its state. Prepared campaign data and live LinkedIn actions are different counts, different permissions and different risks.

Do not confuse a human-looking browser with permission

Some search results teach Claude to operate a logged-in Chrome window, visit profiles, click Connect and send notes with randomized delays. A vendor may call that “safer” because the browser looks less like a script. That is a claim about detection, not about authorization. LinkedIn's automated-activity rule says third-party software and browser extensions that scrape, modify or automate its website are not allowed. It does not publish a small-batch exception for Claude.

The rule can apply even when the note is useful and a person reviews the sequence. Before enabling any browser or site-action integration, examine LinkedIn's current policy, the integration's access to your account and what happens on a warning or restriction. You can still use Claude for research summaries, deduplication, writing and CRM preparation if you decide the site-action risk is unacceptable.

Do not give Claude session cookies, passwords or a broad “send every day” instruction. Do not use a campaign count as evidence that a message was delivered. If an account warning or verification prompt appears, stop the relevant action; it is not an optimization signal to route around.

Check the result in stages

For a small pilot, track distinct stages: prospects reviewed, records saved, messages drafted, campaign prepared, account actions attempted, account actions completed, replies received and conversations qualified. The first four can grow while the last four stay at zero. That is not a broken analytics report; it is an honest picture of what the system has and has not done.

Look at failures before you ask Claude to process more people. A missing field, wrong workspace, revoked key or changed tool availability needs a different fix from a LinkedIn restriction. If a person replies, the follow-up plan should stop or be reviewed. If your connected tool cannot reliably show completion or reply state, do not instruct Claude to calculate a reply rate from drafted messages.

There is no universal “safe number” of invitations to put in this guide. Vendors publish different figures, and LinkedIn's policy does not bless a threshold. The more useful pilot goal is a decision you can audit: did Claude save legitimate review time, keep source evidence attached, and avoid unapproved changes?

Common questions

Can Claude connect directly to LinkedIn?

Claude can use a browser or connected third-party tools, but a browser action is not the same thing as an approved LinkedIn integration. For this workflow, Claude connects to OutreachGenie's workspace via MCP. That gives it the workspace capabilities you authorize, not blanket access to LinkedIn's website.

Do I need Claude Code?

No. Claude.ai and Cowork can use a custom remote connector; Claude Code is useful if your work also involves local files or a CLI. Use the client whose permissions and workflow you can inspect. The connector may be disabled for a conversation even after you added it to your Claude account.

Will connecting OutreachGenie start sending messages?

No. Connecting a tool and saving a prospect are separate from preparing a flow and launching a campaign. Confirm the current campaign state in the product before any activation, and treat any LinkedIn account action as a separate policy decision.

Why isn't Claude finding my list?

Check the chosen OutreachGenie workspace, connector toggle, OAuth scopes or API-key status, and the actual MCP tool list. If a read-only lookup still fails, do not ask Claude to guess from old chat context. Resolve the connection before a write.

Start with OutreachGenie

If you want to make Claude useful for LinkedIn outreach without losing the handoffs, OutreachGenie keeps prospects, source notes, lists, flows and campaigns in one reviewable workspace. Follow the Claude setup guide, connect one workspace, ask for a read-only inspection, then prepare a small list and inactive campaign draft. Approve each change after you can see what it will touch.

OutreachGenie can make the preparation more organized and visible. It cannot cancel LinkedIn's policy or guarantee replies. For the broader explanation of MCP-connected prospecting, see our OutreachGenie MCP guide.