An automated job application pipeline built on Claude Code that scans job boards, scores roles against your profile, generates tailored materials, and submits applications — with human-in-the-loop approval before every submission.
Applying to jobs at scale is broken. You either spray-and-pray with a generic resume, or spend 45 minutes per application tailoring materials manually. This agent does the tailoring at scale while keeping you in control of what actually gets submitted.
Scanner Agent (scheduled)
↓ scans LinkedIn alerts, Indeed, career pages via Gmail + MCP
↓ scores each role 0-10 against your profile
↓ filters: <5 skip, 5-6 track, 7+ high-priority, 8+ auto-generate
↓
Materials Agent (auto-triggered for score 8+)
↓ selects best resume cluster for the role
↓ generates tailored resume + cover letter
↓ uploads to Google Drive
↓
Submitter Agent (manual trigger only)
↓ detects ATS platform (Greenhouse, Lever, etc.)
↓ fills form fields from your candidate profile
↓ STOPS and asks for your approval before submitting
↓ records confirmation in Google Sheet tracker
Not all job sites are created equal for automation:
| Tier | Platforms | Method | Friction |
|---|---|---|---|
| 1 | LinkedIn Easy Apply, Indeed, Dice | MCP / Chrome extension | Lowest — API or simple DOM |
| 2 | Greenhouse, Ashby, Lever | Chrome MCP (DOM automation) | Medium — standardized forms |
| 3 | Oracle HCM, Workday, iCIMS | Manual assist | Highest — native file pickers, CAPTCHAs |
The agent prioritizes Tier 1 applications first, then works through Tier 2, and flags Tier 3 for your manual intervention.
Instead of one resume, maintain 3-5 cluster variants that emphasize different strengths:
{
"clusters": {
"ai_builder": { "emphasis": "production AI systems, technical architecture" },
"ai_advisory": { "emphasis": "C-suite advisory, transformation strategy" },
"ai_partnerships": { "emphasis": "strategic alliances, ecosystem development" },
"product_marketing": { "emphasis": "GTM strategy, product positioning" }
}
}The Materials Agent automatically selects the best cluster for each role based on JD keyword analysis.
- Claude Desktop with Claude Code / Cowork mode
- Node.js 18+
- Google account (for Sheets tracker + Drive storage)
- LinkedIn, Indeed accounts
# 1. Clone and install
git clone https://github.com/nycsav/job-application-agent.git
cd job-application-agent
npm install
# 2. Create your candidate profile
cp config/candidate.example.json config/candidate.json
# Edit candidate.json with your real data (it's gitignored)
# 3. Add your resume PDFs
mkdir -p materials/resumes
# Place your cluster-variant PDFs here (gitignored)
# 4. Configure Google Sheet tracker
# Create a Google Sheet and update the sheet ID in CLAUDE.md
# 5. Run a manual scan
claude "Read routines/daily-scan.md and execute the steps"# Set up automated scanning (8 AM & 5 PM weekdays)
claude routine add job-scanner \
--schedule "0 8,17 * * 1-5" \
--prompt "$(cat routines/daily-scan.md)"job-application-agent/
├── agents/
│ ├── scanner.mjs # Job discovery + scoring
│ ├── materials.mjs # Resume/cover letter generation
│ └── submitter.mjs # Form filling with human gate
├── config/
│ ├── candidate.json # YOUR profile (gitignored)
│ ├── candidate.example.json # Template for new users
│ ├── roles.json # Per-role customizations
│ ├── platforms.json # ATS platform capability matrix
│ └── resume_map.json # Cluster → resume file mapping
├── lib/
│ ├── template-engine.mjs # Template merging
│ └── docx-builder.mjs # ATS-optimized .docx generation
├── templates/
│ └── master_cover_letter.md # Cover letter template
├── routines/
│ ├── daily-scan.md # Scheduled scan routine
│ ├── generate-materials.md # Materials generation routine
│ └── submit-applications.md # Submission routine (manual)
├── materials/
│ ├── resumes/ # Your PDFs (gitignored)
│ └── cover-letters/ # Generated letters (gitignored)
└── apps-script/
└── SCORE_FILTER_PATCH.md # Google Apps Script integration
Each role is scored 0-10 across 5 dimensions:
| Dimension | Max Points | Signals |
|---|---|---|
| Title match | 3 | Director/VP/Head = 3, Manager/Lead = 2, Other = 1 |
| Skill overlap | 3 | Keyword matches against your skill list |
| Industry fit | 2 | Strong industry = 2, Adjacent = 1, Unrelated = 0 |
| Location | 1 | Preferred city or Remote = 1 |
| Compensation | 1 | Within target range = 1 |
- Never auto-submits — human approval required for every application
- Exclusion list — companies you handle personally or want to avoid
- No fabrication — only uses verified metrics from your candidate profile
- Screenshot audit trail — captures form state before submission
- Deduplication — checks Google Sheet before applying to prevent duplicates
This project is built on Claude Code and uses several agentic patterns worth highlighting:
The pipeline uses multiple MCP servers simultaneously — each one gives Claude direct API access to a service without browser automation:
- Gmail MCP — scans LinkedIn/Indeed alert emails, extracts job details, applies processed labels
- Google Sheets MCP — reads/writes the job tracker (scores, statuses, Drive links)
- Google Drive MCP — uploads tailored resumes and cover letters directly
- Indeed MCP — searches jobs and pulls full JD details via API
- Chrome MCP — DOM-aware browser automation for ATS form filling (Greenhouse, Lever, etc.)
The key insight: MCP connectors are dramatically faster and more reliable than browser automation. A Gmail search via MCP takes ~1 second; the same search via browser automation takes 15-20 seconds with scrolling, clicking, and screenshot parsing. We use browser automation only when no MCP exists (LinkedIn) or when the ATS requires visual interaction.
The current pipeline runs sequentially: scan → score → generate → submit. The architecture supports parallelism at two levels:
-
Batch scoring — score multiple roles simultaneously using Claude Code's Task tool to spawn parallel subagents. Each subagent gets one role + the candidate profile and returns a score independently.
-
Parallel materials generation — generate tailored resumes for multiple score-8+ roles concurrently. Each subagent selects the appropriate resume cluster, generates the cover letter, and uploads to Drive.
-
Multi-platform submission — submit to multiple Tier 1 platforms (LinkedIn Easy Apply, Indeed) in parallel, with a shared deduplication check against the Google Sheet.
Claude Code Routines run the scanner on a cron schedule:
# Scan at 8 AM and 5 PM weekdays
claude routine add job-scanner \
--schedule "0 8,17 * * 1-5" \
--prompt "$(cat routines/daily-scan.md)"Each routine execution is a fresh Claude session that reads the current config, scans all sources, scores new roles, and sends an email summary — no persistent process needed.
The submitter agent reads config/platforms.json to determine HOW to submit before attempting it. This avoids wasting cycles on sites that will block automation:
Tier 1 (MCP/API): LinkedIn Easy Apply, Indeed, Dice
Tier 2 (Chrome MCP): Greenhouse, Ashby, Lever
Tier 3 (Manual): Oracle HCM, Workday, iCIMS
This tiering was discovered empirically — Oracle HCM (used by JPMorgan, many Fortune 500) blocks programmatic file uploads entirely, opening a native OS file picker that no browser automation can interact with.
- Claude Code — agent orchestration and MCP integration
- Model Context Protocol (MCP) — direct API access to Gmail, Google Sheets, Google Drive, Indeed
- Claude in Chrome — DOM-aware browser automation for ATS form filling
- Google Sheets API — application tracking
- Google Drive API — materials storage
- docx — ATS-optimized document generation
This project is designed to help job seekers automate the tedious parts of job hunting while keeping humans in control. PRs welcome — especially for:
- New ATS platform support in
config/platforms.json - Improved scoring heuristics
- Additional resume cluster strategies
- Integration with more job boards
MIT