MCP-Linkedin AGENTS.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Development Commands

• Use uv for dependency management: uv sync (dev: uv sync --group dev)
• Lint: uv run ruff check . (auto-fix with --fix)
• Format: uv run ruff format .
• Type check: uv run ty check (using ty, not mypy)
• Tests: uv run pytest (with coverage: uv run pytest --cov)
• Pre-commit: uv run pre-commit install then uv run pre-commit run --all-files
• Run server locally: uv run -m linkedin_mcp_server --no-headless
• Run via uvx (PyPI/package verification only): uvx mcp-server-linkedin
• Docker build: docker build -t linkedin-mcp-server .
• Install browser: uv run patchright install chromium

Scraping Rules

• One section = one navigation. Each entry in PERSON_SECTIONS / COMPANY_SECTIONS (scraping/fields.py) maps to exactly one page navigation. Never combine multiple URLs behind a single section.
• Minimize DOM dependence. Prefer innerText and URL navigation over DOM selectors. When DOM access is unavoidable, use minimal generic selectors (a[href*="/jobs/view/"]) — never class names tied to LinkedIn's layout.
• Detection must be locale-independent. Classification logic — connection state, action availability, button identity — must rely on URL patterns (/preload/custom-invite/?vanityName=USER, /in/USER/edit/intro/, /messaging/compose/), attribute presence (aria-label exists, aria-expanded exists, aria-disabled exists), or structural counts — never on text values like "Connect", "Follow", "Message", "1st", "Pending". The verb in an aria-label is locale-dependent; whether the attribute exists is not. Where text is genuinely the only signal, guard it behind an explicit per-locale table and document the limitation in code.

Browser Identity Rules

• The browser must not contradict itself. Anything it says about itself has to survive being checked against another surface of the same browser: the user-agent against sec-ch-ua, the page against its workers and iframes, the reported screen against the window sitting on it. The goal is coherence, not invisibility — invisibility cannot be proven, while a contradiction is a fact and can be found by anyone who looks twice.
• Never inject a fingerprint. No user_agent, no custom headers, no spoofed client hints. Patchright's own guidance says the same, and every override measured here made things worse: a user_agent= argument changes the string but not the client hints, and never reaches service workers at all (playwright#5237, closed as an upstream Chromium bug). A browser telling the truth beats one caught lying.
• A proxy must contain every egress path, not just HTTP. WebRTC uses UDP and went around a configured proxy until the switches in browser_launch.py. DNS and QUIC belong to the same family; check them before assuming a new setting is contained.
• Verify identity changes by measurement. docs/browser-fingerprint.md lists the four detectors, what each one alone would miss, and the values measured so far. A launch-configuration change without a measurement against them is a guess.