autopilot-jobhunt 0.4.0

AI-powered job scanner, scorer, and application drafter. Finds jobs while you sleep.

autopilot-jobhunt

Your AI job agent. Finds, scores, and drafts applications — while you sleep.

Scans 130+ company careers pages nightly → scores every role against your resume with an LLM → sends you the top matches on Telegram → drafts a tailored resume + cover letter on demand.

📖 Full setup guide with Claude Code MCP integration → SETUP.md

---

How it works

flowchart LR
    A["🌐 130+ Careers Pages"] -->|TinyFish API| B["Job Discovery"]
    B --> C["LLM Batch Scorer\n(0–100 fit score)"]
    C -->|score ≥ min| D["📱 Telegram Alert\nTop N matches"]
    C -->|on demand| E["✉️ Cover Letter\n+ Resume Bullets"]
    C --> F["📊 CSV Export"]

The scoring prompt uses your actual resume — not keywords. The LLM reads your full work history and the job description, then explains in one sentence why you fit or don't. No more guessing.

What a scan result looks like

Scanning Mistral AI...
  3 new job URLs. Fetching details...
  Scoring jobs...
  Saved 2 jobs from Mistral AI

Scanning HuggingFace...
  5 new job URLs. Fetching details...
  Scoring jobs...
  Saved 3 jobs from HuggingFace

Scanning Stripe...
  No new jobs found
...
Scan complete.
Top 5 sent to Telegram.

What the Telegram notification looks like

Job Hunt — 06 Jun 2026
5 matches found

#1 | Mistral AI | Applied AI Engineer, ML Infrastructure
📍 Paris/London/Marseille, On-site
🔧 Python, LLMs, RAG, AWS, MLOps, DevOps
✅ Role combines applied AI + ML infrastructure in EU, aligns with MLOps/RAG expertise and relocation goal
Score: 85/100  →  https://jobs.lever.co/mistral/...

#2 | HuggingFace | Staff ML Engineer
📍 Remote (EU)
🔧 Python, PyTorch, Transformers, CUDA, MLOps
✅ Open-source ML role matches deep learning and distributed training background
Score: 80/100  →  https://apply.workable.com/huggingface/...

...

Reply "apply to #N" to draft a tailored application.

---

What it does

Every night at 2:30 AM:
  ┌─────────────────────────────────────────────────────────┐
  │  Scans careers pages  →  Scores with LLM  →  Notifies  │
  │       (130+ cos)           (0–100 fit)       (Telegram) │
  └─────────────────────────────────────────────────────────┘

On demand:
  autopilot draft 1  →  tailored resume + cover letter in 60s

---

Usage modes

Mode 1: Standalone CLI (no Claude Code required)
  pip install autopilot-jobhunt
  autopilot scan / autopilot draft 1 / autopilot export

Mode 2: Claude Code MCP (control via natural language)
  pip install 'autopilot-jobhunt[mcp]'
  claude mcp add autopilot-jobhunt ...
  → "Scan for ML jobs" / "Draft application for job #2"

Both modes use the same config and produce the same output.

Quick start

Option A — pip install

pip install autopilot-jobhunt        # or: pip install 'autopilot-jobhunt[mcp]' for Claude Code
mkdir my-job-hunt && cd my-job-hunt
autopilot init                       # creates config.json, companies.json, resume/, .env
# Fill in config.json (API keys + your profile) and resume/YOUR_RESUME.md, then:
autopilot scan

Option B — clone (recommended if you want to customize companies or contribute)

git clone https://github.com/tarunlnmiit/autopilot-jobhunt.git
cd autopilot-jobhunt
pip install -e '.'               # standalone CLI
# pip install -e '.[mcp]'       # + Claude Code MCP integration
cp config.example.json config.json && cp .env.example .env
# Fill in your API keys and candidate profile, then:
autopilot scan

For the full walkthrough — API key setup, Claude Code MCP registration, rate limit details, and troubleshooting — see SETUP.md.

API keys needed

Service	Cost	Required	Where to get it
TinyFish	Free — no credit card	Always	agent.tinyfish.ai
OpenRouter	Free — 4-model fallback chain	Unless using Claude CLI / Anthropic	openrouter.ai
Telegram	Free	Optional	@BotFather on Telegram

---

Claude Code / MCP integration

Use autopilot-jobhunt as an MCP server inside Claude Code (CLI) or Claude Desktop.

Step 1: Install with MCP support

git clone https://github.com/tarunlnmiit/autopilot-jobhunt.git
cd autopilot-jobhunt
pip install -e '.[mcp]'

Step 2: Register with Claude Code

Option A — one command:

claude mcp add autopilot-jobhunt \
  --env TINYFISH_API_KEY=your_key \
  --env OPENROUTER_API_KEY=your_key \
  --env TELEGRAM_TOKEN=your_token \
  --env TELEGRAM_CHAT_ID=your_chat_id \
  -- python -m job_hunt.mcp_server

Option B — edit ~/.claude.json manually:

{
  "mcpServers": {
    "autopilot-jobhunt": {
      "command": "python",
      "args": ["-m", "job_hunt.mcp_server"],
      "cwd": "/absolute/path/to/autopilot-jobhunt",
      "env": {
        "TINYFISH_API_KEY": "your_key",
        "OPENROUTER_API_KEY": "your_key",
        "TELEGRAM_TOKEN": "your_token",
        "TELEGRAM_CHAT_ID": "your_chat_id"
      }
    }
  }
}

Note: cwd must point to the cloned repo — the server reads config.json and companies.json from there.

Step 3: Use it

In any Claude Code session:

"Scan for ML jobs"
"Draft an application for job #2"
"Export jobs from the last 7 days with score above 70"

Claude Desktop

Same JSON block — add it under mcpServers in Claude Desktop → Settings → Developer.

---

Customize your target companies

Edit companies.json. Each entry needs:

{
  "name": "Stripe",
  "careers_url": "https://stripe.com/jobs",
  "search_domain": "stripe.com",
  "location": "Remote / San Francisco, CA",
  "region": "Remote"
}

The repo ships with 130+ pre-configured EU, NZ, and remote-friendly tech companies. Add or remove as you like.

---

How scoring works

The LLM reads your full resume + the full job description and assigns a score 0–100:

Score	Meaning
80–100	Near-perfect fit — apply immediately
60–79	Good fit — worth applying
40–59	Partial fit — apply if pipeline is thin
< 40	Poor fit — skipped

Set min_score in config to filter. Default: 60.

---

Project structure

autopilot-jobhunt/
├── job_hunt/
│   ├── main.py          # CLI entry point
│   ├── scanner.py       # Job discovery + LLM scoring
│   ├── drafter.py       # Resume tailoring + cover letter
│   ├── notifier.py      # Telegram notifications
│   ├── llm_utils.py     # OpenRouter wrapper with fallback
│   ├── tools.py         # Protocol-agnostic tool layer
│   └── mcp_server.py    # MCP server (Claude/AI assistant integration)
├── demo/                # Demo scripts for recording GIF
├── resume/              # Put your resume here (gitignored)
├── state/               # Scan state (gitignored)
├── output/              # Generated applications (gitignored)
├── companies.json       # 130+ target companies
├── config.example.json  # Config template (copy to config.json — gitignored)
└── config.json          # Your config (gitignored — never committed)

---

LLM options

Default: OpenRouter (free)

Uses a 4-model fallback chain — all free, no credit card needed:

Model	Role
meta-llama/llama-3.3-70b-instruct:free	Primary — best quality
nvidia/nemotron-3-super-120b-a12b:free	Fallback 1 — 120B
google/gemma-4-31b-it:free	Fallback 2
qwen/qwen3-coder:free	Fallback 3

If one model hits its daily free-tier quota, the tool automatically tries the next. Zero LLM cost by default.

Alternative A: Claude Code CLI (no API key needed)

If you have Claude Code installed and authenticated, you can use it as the LLM backend — no separate API key required:

In config.json:

"llm_provider": "claude_cli"

Or via environment variable: LLM_PROVIDER=claude_cli autopilot scan

Optionally set a model: "claude_cli_model": "sonnet" (or "opus", "haiku", empty = Claude's default).

Note: Requires the claude binary in your PATH. Verify with claude --print "hi" first. The MCP server and cron jobs must run in an environment where your claude auth session is active.

Rate-limit note: Each call loads your global Claude Code context (~25–30k tokens). A nightly scan (5–15 LLM calls) burns significantly against your subscription's 7-day rate limit. Prefer OpenRouter for nightly automation; use Claude CLI for occasional on-demand drafts.

Alternative B: Anthropic API

If you have an Anthropic API key:

pip install 'autopilot-jobhunt[claude]'

In config.json:

"llm_provider": "anthropic",
"anthropic_api_key": "sk-ant-...",
"anthropic_model": "claude-haiku-4-5-20251001"

claude-haiku-4-5-20251001 is fast and cheap; claude-sonnet-4-6 gives higher quality scores. A nightly scan uses ~5–15 LLM calls total (jobs scored in batches of 10).

---

Contributing

See CONTRIBUTING.md. PRs welcome for:

• Adding companies to companies.json
• New ATS platform support (Rippling, Lever variants, Workday)
• OpenAI / Gemini MCP adapters
• Better scoring prompts

---

License

MIT — see LICENSE.

---

Built by @tarunlnmiit. If this saved you hours of job searching, a ⭐ means a lot.