What this actually is
Model Context Protocol is the standard Claude uses to talk to outside systems. An MCP server exposes a set of tools, Claude reads what they do, and then it can call them during a conversation. No plugin store, no browser extension, no glue code.
There is no MCP server from LinkedIn and there will not be one. LinkedIn’s user agreement prohibits unauthorised automated access to the platform, and anyone selling you a direct Claude-to-LinkedIn bridge is selling you a restriction.
What works instead is one step removed. Prospectio holds the LinkedIn session, enforces the sending limits and routes through residential IPs. The MCP server gives Claude a way to drive Prospectio. Claude never touches LinkedIn directly.
Setting it up
Get an API key first: in Prospectio, go to Settings → Integrations and generate one.
Local install. Add this to your Claude config:
{
"mcpServers": {
"prospectio": {
"command": "npx",
"args": ["-y", "salesrobot-cli"],
"env": {
"SR_API_KEY": "your_api_key_here"
}
}
}
}Hosted, no install. If you would rather not run anything locally:
{
"mcpServers": {
"prospectio": {
"url": "https://mcp.salesrobot.co",
"headers": {
"Authorization": "Bearer your_api_key_here"
}
}
}
}Clients that support OAuth with PKCE can point at the same URL and auto-discover the metadata at /.well-known/oauth-authorization-server rather than handling a key at all.
On the package name: Prospectio runs on infrastructure shared with SalesRobot, so the npm package is published as salesrobot-cli. Your Prospectio key works with it. Better you read that here than wonder about it mid-install.
If npx picks the wrong Node version, point command at the full path to your Node binary instead. That is the single most common setup failure and it has nothing to do with the server.
What Claude can do once it’s connected
Thirty-three tools. The ones that matter day to day:
Building a campaign. campaign_create, campaign_update_settings, campaign_start, campaign_pause, campaign_stats, campaign_list.
Getting prospects in. This is the part that saves the most time. prospect_import_linkedin_search pulls from a Sales Navigator search. prospect_import_linkedin_group and prospect_import_linkedin_event do what they say. prospect_import_post_reactions and prospect_import_post_comments pull the people who engaged with a specific post, which is the highest-intent list available on LinkedIn and a pain to build by hand. Plus prospect_upload_csv, prospect_sync_sales_nav and prospect_import_recruiter_lite.
Running the inbox. inbox_get_messages, inbox_send_message, inbox_tag_chat.
Housekeeping. linkedin_account_list, linkedin_account_check, prospect_pause, prospect_update, ai_features_create_variable.
The order matters. Account, then campaign, then prospects, then settings, then sequence, then start. Claude will work it out from the tool descriptions, but if you ask it to start a campaign before there are prospects in it you will get an error rather than magic.
What to actually ask it
The setup is not the interesting part. This is:
Pull everyone who commented on [post URL], filter to heads of engineering at companies over 200 people, and show me the list before you add anyone to a campaign.
Create a campaign called Q4 Engineering, add these prospects, and write a first message that names their own tooling and ends in a question. Keep it between 240 and 420 characters and do not mention what we sell.
Show me campaign stats for the last two weeks and tell me which sequences are below a 30 percent reply rate.
Read my inbox, list anything that came in since yesterday and hasn’t been answered, and draft replies. Do not send them.
That last one is worth more than the rest combined. Across 10,120 inbound replies in our dataset, the ones answered inside 24 hours converted to a meeting at 21.4 percent. The ones answered after 48 hours converted at 3.6 percent. Nearly six times the difference, on replies you have already paid to generate. An agent that reads the inbox every morning and drafts responses closes exactly that gap.
Keep the approval step
The tools can send. That does not mean you should let them, at least not at first.
Ask for drafts, read them, then send. Our copy tests are unambiguous about what happens when nobody is checking: a message with solution detail in it replied at 18.6 percent, and the same message cut to two and a half sentences with no pitch replied at 44.5 percent. Left alone, a model will write the longer one, because it reads as more helpful and it is wrong.
Two other things worth building into your instructions. Ask for 240 to 420 characters, which peaked at 38.6 percent replies against 11.2 percent for anything over 700. And tell it to name the prospect’s own product or a direct competitor, which took reply rate from 17.1 to 34.5 percent across 10,815 prospects.
None of that is Claude being bad at writing. It is Claude not knowing your constraints until you give them to it, which is the whole argument for encoding them properly.
Where this fits
The honest framing is that connecting Claude to your outreach platform does not make the outreach better. It makes it faster to set up and faster to respond, and one of those two is worth a great deal.
The three largest effects in our dataset are verified subject-matter targeting at 4x, personalised video in the first touch at plus 40 percent, and the 24-hour reply window at 6x. An agent with these tools can genuinely help with the first and the third. It cannot record your video for you.
Full method and all fifteen controlled tests are on the 2026 outbound benchmarks page, free and ungated.