linkedin-mcp-server

Exposes LinkedIn person profiles as MCP tools callable by Claude and other MCP-compatible AI assistants. Uses Patchright (a hardened Playwright fork) to maintain persistent browser profiles stored locally (~/.linkedin-mcp/profile) with cookie-based authentication, eliminating repeated login flows. Implements a 'one-section-one-navigation' architecture where each profile section (work history, education, skills, certifications, posts) maps to a discrete URL, allowing the AI to request only needed data and minimize page loads.

Retrieves comprehensive company data from LinkedIn including overview, employees, recent feed posts, and company metadata through MCP tools. Implements the same 'one-section-one-navigation' pattern as person profiles, where each company section (overview, employees, feed) maps to a specific URL. Uses Patchright browser automation to parse company pages and extract structured data without triggering rate limits or detection.

Provides Docker and docker-compose configurations for containerized deployment of the LinkedIn MCP server. Enables users to run the server in isolated containers with predefined dependencies, environment variables, and volume mounts for profile persistence. Supports both standalone Docker runs and multi-container orchestration via docker-compose, simplifying deployment across different environments (local, cloud, CI/CD).

Integrates with Claude Desktop through a manifest.json file that registers the LinkedIn MCP server as a tool provider. The manifest defines tool schemas (input parameters, output types) and server connection details, enabling Claude Desktop to discover and invoke LinkedIn tools. Uses Claude Desktop's native MCP client to communicate with the server via stdio or network sockets.

Implements a 'one-section-one-navigation' design pattern where each data section (person work history, company overview, job details) maps to exactly one URL. This allows Claude to request only the specific data it needs without loading entire profiles or pages. Reduces page loads, minimizes rate-limit exposure, and improves reliability by limiting the DOM parsing surface area. Each tool corresponds to a discrete navigation action, enabling granular data fetching.

Enables Claude to search LinkedIn job listings with filters (keywords, location, experience level, job type, salary range) and retrieve detailed job information by ID. Implements structured search parameters that map to LinkedIn's search API query format, allowing the AI to construct filtered job searches without manual URL manipulation. Returns job metadata including title, company, location, salary, description, and application requirements.

Retrieves LinkedIn inbox conversations and enables message search across threads. Implements conversation listing (fetching recent inbox threads) and message search (finding specific messages within conversations). Uses Patchright to navigate LinkedIn's messaging interface and extract conversation metadata (participants, timestamps, message content). Maintains conversation threading context for multi-turn message analysis.

Enables Claude to send LinkedIn connection requests programmatically, optionally including personalized messages. Implements form submission via Patchright to navigate LinkedIn's connection request flow, including message composition and submission. Handles LinkedIn's rate limiting and connection request validation (e.g., preventing duplicate requests to the same person).

Implements a 'Sequential Middleware Lock' that serializes all MCP tool calls within a single process to prevent LinkedIn from detecting multiple concurrent sessions from the same account. Uses a queue-based execution model where simultaneous tool invocations are queued and executed serially. This design decision protects against account suspension and detection while maintaining a single authenticated browser session across all tools.

Manages persistent browser profiles stored locally (default: ~/.linkedin-mcp/profile) that maintain LinkedIn authentication state (cookies, session tokens) across server restarts. Eliminates repeated login flows by reusing authenticated browser contexts. Implements profile initialization on first run (prompting for credentials) and automatic session refresh on subsequent runs. Uses Patchright to manage browser lifecycle and profile persistence.

Implements the Model Context Protocol (MCP) server using the FastMCP framework, which handles MCP protocol serialization, tool registration, and request/response lifecycle management. Exposes all LinkedIn scraping tools as MCP-compliant tools callable by Claude Desktop and other MCP-compatible AI assistants. Manages tool schema definition (input parameters, output types) and error handling within the MCP protocol contract.

Uses Patchright, a hardened fork of Playwright designed specifically for automation evasion, to bypass LinkedIn's advanced bot detection mechanisms. Patchright patches Playwright's WebDriver exposure, randomizes browser fingerprints, and implements anti-detection heuristics that prevent LinkedIn from identifying automated browser sessions. This enables reliable scraping without triggering account restrictions or CAPTCHAs.

Manages server configuration through environment variables (.env file) and credential storage, allowing users to customize LinkedIn credentials, profile directory, and server behavior without code changes. Implements credential validation on startup and provides clear error messages for missing or invalid configuration. Supports both interactive credential prompts (first run) and non-interactive credential loading (subsequent runs).