LinkedIn research and outreach agent built on the LinkedIn MCP server with Codex

What they did

They wanted an agent that could do the boring first pass of LinkedIn research without copy-pasting between tabs. Using the open-source linkedin-mcp-server, they exposed LinkedIn actions to their agent through their own logged-in browser session, then drove it from Codex. The agent was pointed at a shortlist of target companies and people: it pulled each profile, summarized role, tenure, and recent activity, pulled company pages for headcount and positioning, and checked open job postings where relevant. From those reads it produced a compact research brief per prospect and drafted a short, specific connection note or follow-up message referencing something real from the profile. They also used it to triage the inbox — grouping conversations and flagging ones that needed a reply — and to scan job listings for a separate search. Everything ran locally against their own session rather than a scraped dataset, so the agent saw what they would see when logged in.

How it went

The research pass that used to take an afternoon per batch collapsed into a single agent run with a review step. Briefs were consistent in structure and the drafted notes needed light editing rather than rewriting. The inbox triage was the surprise win — it surfaced stale threads they had forgotten. Polish level is solid for internal use; nothing was auto-sent, every message went through human review before going out.

Worth knowing

This rides on your personal LinkedIn session, so rate limits, session expiry, and account-risk policies all apply — keep volumes low and human in the loop. The project is independent and not affiliated with LinkedIn, and the README leans on sponsored hosted alternatives, so read the setup docs carefully. Scraping-adjacent automation can break when LinkedIn changes its UI, and you should not use it to mass-message or automate connections at scale.

Wiring

Install via uvx (or the Claude Desktop MCPB / Codex plugin path) so the server runs locally and authenticates through your own browser session. Register the server as an MCP tool source for Codex, confirm the profile/company/messaging tools appear, then keep credentials and session cookies out of the repo. Order of operations: authenticate once, smoke-test a single profile read, then run the batch prompt with a hard cap on reads. Store targets and outputs as plain files in the repo so runs are diffable, and gate any send action behind manual approval rather than exposing it to the agent.

Try it yourself

Set up the linkedin-mcp-server locally and connect it to Codex. Then, for the list of target companies and people in targets.md: (1) pull each person's profile and write a 5-line brief covering current role, tenure, prior companies, and one recent activity worth referencing; (2) pull each company page and note headcount, positioning, and any open roles relevant to my search; (3) draft a connection note under 300 characters that references something specific and non-creepy from the profile; (4) list my unread LinkedIn conversations and flag any that have gone unanswered for more than 7 days. Output everything to research/briefs.md and drafts/notes.md. Do not send anything — drafts only, and keep total profile reads under 40 for this run.