mcp-job-search

An MCP server that searches Berlin job listings across Arbeitnow, Stepstone, LinkedIn, Indeed, and Glassdoor, filters for English-friendly roles, and builds assisted-apply packets — cover letter, outreach email, LinkedIn referrer message, and apply URL — for each role.

It does not auto-submit applications. You stay the human pressing send.

Install

The fastest path uses uv (one command, no virtualenv juggling):

curl -LsSf https://astral.sh/uv/install.sh | sh   # if you don't have uv

Then add this block to your Claude Desktop config:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "job-search": {
      "command": "uvx",
      "args": ["--from", "mcp-job-search", "mcp-job-search"]
    }
  }
}

Restart Claude Desktop. uvx downloads and caches the package on first call.

Tools

Slash-command prompts

How to use it

The fastest path is the slash command:

/start_job_search

Claude walks you through CV path, email, category, and location, then searches and surfaces the top 5 results grouped by company. You pick which to expand into an application packet.

Or just talk in plain English:

"Find me senior Python backend roles in Berlin, English-only."

"Build an application packet for job arbeitnow:backend-engineer-acme."

"Rewrite my CV bullets for the role at N26."

Configuration

The server reads ~/.config/mcp-job-search/config.yaml if present, else the bundled defaults (Berlin, English-only). For per-session overrides — your resume, your email, your name — use the set_resume and set_candidate tools (Claude calls them automatically during the guided flow).

If you want defaults to persist across sessions, drop a config file at the path above:

location: "Berlin"
english_only: true
resume_path: "/Users/you/Documents/resume.pdf"
candidate:
  name: "Your Name"
  email: "you@example.com"
  linkedin_url: "https://linkedin.com/in/yourhandle"
  pitch: "Backend engineer with 5+ years building Python services"

Limits + honest caveats

• Arbeitnow: free public API, no key. Coverage skews toward Berlin/EU tech and remote roles. The API has no server-side search, so the client paginates a few pages and filters locally.
• Stepstone: HTML scrape. Brittle when they redesign — open an issue if results start returning empty.
• JobSpy (LinkedIn / Indeed / Glassdoor): also scrape-based, via the python-jobspy package. LinkedIn rate-limits aggressively and Glassdoor's location parser sometimes errors on German city strings — Indeed is the most reliable of the three.
• English filter: heuristic. A few German-language postings that mention "Englisch" will slip through, and a few English-friendly ones with mostly German boilerplate will get cut. The min_results top-up will fill the gap with German jobs (tagged language='de') so you always get a list.
• Cover letters / referral messages: deterministic templates. Claude refines per job — you copy out.

Dev install (contributing)

git clone https://github.com/heykay/mcp-job-search
cd mcp-job-search
uv sync
uv run pytest

Local Claude Desktop entry pointing at your clone:

"job-search": {
  "command": "uv",
  "args": ["--directory", "/absolute/path/to/mcp-job-search", "run", "mcp-job-search"]
}

License

MIT — see LICENSE.