JobScout MCP

An MCP (Model Context Protocol) server that lets Claude search live job postings, score them against your own skills, and track which ones you've applied to - all running locally on your machine.

Why this exists

Built as a hands-on project to learn how MCP servers actually work: how Claude discovers tools, how tool schemas get defined, and how to design tools that are genuinely useful rather than just API wrappers.

Related MCP server: job-search

Directory structure

jobscout-mcp/
├── server.py              # MCP server entrypoint - defines all 4 tools
├── scoring.py              # Scores a job posting against profile.json
├── storage.py              # SQLite-backed tracking of application status
├── profile.json             # YOUR editable skills/preferred-roles list
├── requirements.txt         # Runtime dependencies
├── requirements-dev.txt      # Adds pytest for running tests
├── .gitignore
├── LICENSE
├── README.md
└── tests/
    ├── test_scoring.py       # Unit tests for scoring.py
    └── test_storage.py       # Unit tests for storage.py

How the pieces connect: server.py is the only file that talks to Claude - it imports functions from scoring.py and storage.py and exposes them as tools. scoring.py reads profile.json (your skills) and never touches the network or the database. storage.py owns the local SQLite file (jobscout.db, created automatically on first run, gitignored so it stays private) and never touches the network either. This separation is why tests/ can test scoring and storage logic directly, without needing a live network connection or a running MCP server.

Tools exposed to Claude

Setup

git clone <your-repo-url>
cd jobscout-mcp
python3 -m venv venv
source venv/bin/activate       # on Windows: venv\Scripts\activate
pip install -r requirements.txt

Edit profile.json with your actual skills and preferred role keywords before using rank_jobs_for_me - the scoring is only as good as this list.

Running the tests

pip install -r requirements-dev.txt
pytest tests/ -v

All storage tests run against temporary throwaway databases (via pytest's tmp_path fixture), so they never touch your real jobscout.db.

Quick manual test (no Claude needed)

python3 -c "
import asyncio
from server import search_jobs, rank_jobs_for_me
print(asyncio.run(search_jobs('python', limit=5)))
print(asyncio.run(rank_jobs_for_me(limit=5)))
"

python3 -c "
from server import track_job, list_tracked_jobs
print(track_job('abc123', 'applied', title='Backend Engineer', company='Acme'))
print(list_tracked_jobs())
"

These hit the live RemoteOK API and a real local SQLite file, so you need actual internet access and a writable directory for this to work.

Connecting to Claude Desktop

Add this to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on Mac, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "jobscout": {
      "command": "python3",
      "args": ["/absolute/path/to/jobscout-mcp/server.py"]
    }
  }
}

Restart Claude Desktop, then try:

• "Use jobscout to find remote backend Python jobs."

• "Rank the live postings against my skills profile."

• "Mark job [id] as applied, I just submitted it."

• "What jobs am I currently tracking as interviewing?"

Known limitations

• Only pulls from RemoteOK's public API - a real production version would pull from multiple job boards.

• Scoring is keyword-based, not semantic - it won't know that "Golang" means the same thing as "go" unless both are in profile.json.

• No automated monitoring of scoring quality over time; see tests/test_scoring.py for what's covered today.

Roadmap (built in this order)

• Part 1: live job search tool

• Part 2: score/rank postings against your own skills

• Part 3: local storage to track seen/applied jobs

• Part 4: tests, docs, packaging