Israeli Multi-Source Tech Job FastMCP Server (TechJobMCP)

An enterprise-grade, privacy-first FastMCP server providing intelligent, multi-source tech job aggregation, smart deduplication, dynamic CV skill & target role extraction, requirement coverage scoring, zero-guesswork Universal DOM form automation, and autonomous job scouting workflows across 11 Israeli & Global sources: Comeet ATS, Greenhouse, Lever, Workday Enterprise, Eightfold.ai, DirectTech (Google, Apple, Amazon, IBM), LinkedIn (Easy Apply & Guest Search), HireMeTech, AllJobs, GotFriends, and Jobify.

📖 Documentation Guides

• 🚀 Setup & Candidate Profile Configuration Guide — Dynamic CV extraction, System 1 LAYA setup, System 2 auto-tailoring, Docker & Python local run.

• 🌐 Cloudflare Tunnel & Port Export Guide — Exposing port 8000 via cloudflared for remote AI clients.

• 🤖 AI Client Integrations Guide — Complete setup for Gemini Spark (autonomous scheduled scout with receipts), Claude Desktop / CoWork, ChatGPT / Codex, and Cursor / Antigravity.

• 🛠️ Spark Prompts & Skills — System prompt and skill instructions for autonomous scouting agents with 11 source badges and audit receipts.

• Contributing — uv-based development, provider guidance, and review expectations.

• Releasing and Changelog — version and manual release policy (0.1.0 is a development version, not a published release).

Related MCP server: MCP Job Application Tracker & Resume Customizer

Quickstart (offline first use)

Requires uv and Python 3.12. From a clean clone:

git clone https://github.com/zvieli/TechJobMCP.git
cd TechJobMCP
uv sync --locked --python 3.12
uv run python -m job_mcp doctor
uv run python -m job_mcp sources

doctor reports required failures separately from optional model, browser, and evaluation-artifact warnings. sources lists the configured source families from the registry without live health checks. Both commands support --json. No .env, local model, Chromium, CV, provider credentials, or application submission is needed for this core read-only first use.

For an MCP-facing metadata operation, list_configured_sources uses the same offline inventory. The existing list_job_sources still performs live health checks. Starting the full MCP server also schedules provider-cache and model warmups, so an actual MCP session is not promised to be offline. Doctor success does not prove provider connectivity, authenticated access, model-backed ranking quality, browser automation, or safe application submission. See setup for optional workflows and reproducibility policy for artifact ownership.

🏛️ Architecture Overview

graph TD
    Client([MCP Client: Gemini Spark / Claude / Cursor / ChatGPT]) --> Tools[FastMCP Server Layer (16 Tools)]
    Tools --> Aggregator[JobAggregator]
    Aggregator --> Registry[SourceRegistry]

    subgraph Parallel Pluggable Sources Layer (11 Sources)
        Registry --> S1[HireMeTechSource<br/>Direct REST API + DOM Fallback]
        Registry --> S2[LinkedInSource<br/>Guest Search + Easy Apply Engine]
        Registry --> S3[ComeetSource<br/>Direct ATS API + iFrame Automation]
        Registry --> S4[GreenhouseSource<br/>Public Boards API + DOM Solver]
        Registry --> S5[LeverSource<br/>Direct API + Structured Postings]
        Registry --> S6[WorkdaySource<br/>Enterprise Workday CXS Direct API]
        Registry --> S7[EightfoldSource<br/>PCSX Search API]
        Registry --> S8[DirectTechSource<br/>Google, Apple, Amazon, IBM Feeds]
        Registry --> S9[JobifySource<br/>Israel Regional Tech Aggregator]
        Registry --> S10[GotFriendsSource<br/>Direct Agency Feed]
        Registry --> S11[AllJobsSource<br/>Portal API & Feed]
    end

    subgraph Processing & Normalization Engine
        S1 & S2 & S3 & S4 & S5 & S6 & S7 & S8 & S9 & S10 & S11 --> Dedup[Deduplication & Entity Merger]
        Dedup --> NormKey["Key = slug(title) + '@' + slug(company)"]
        NormKey --> Merge[Metadata & Links Merger]
    end

    subgraph Dynamic Candidate Engine
        CV["Candidate CV (.pdf / .docx / .txt)"] --> Extractor[Dynamic CV & Profile Extractor]
        Extractor --> Skills["Extracted Skills (40+ tokens)"]
        Extractor --> Stack["Primary Tech Stack (Top Skills)"]
        Extractor --> Seniority["Inferred Seniority & Exclusions"]
        Extractor --> Roles["Dynamic Target Roles"]
    end

    subgraph System 1: Neural Triage Engine (Sub-Millisecond)
        Merge --> Triage{"System 1 Gate"}
        Extractor --> Triage
        Triage --> LayaFT["Fine-Tuned LAYA Engine<br/>Llama-3.2-3B (<1ms Inference)"]
        Triage --> LayaBase["Base LAYA Engine<br/>Zero-Shot Evaluation"]
        Triage --> Heuristic["Deterministic Fallback<br/>Rule-Based Scoring"]
        LayaFT & LayaBase & Heuristic --> SeniorityCap["Seniority & Location Filter<br/>Disqualify or Score (0-100)"]
    end

    subgraph System 2: Cognitive Tailoring Engine
        SeniorityCap --> HighMatch{"Strong Match?<br/>Score ≥ 70/85"}
        HighMatch -- Yes --> Sys2["ApplicationTailoringEngine"]
        Sys2 --> TailoredCV["Custom CV Highlights"]
        Sys2 --> CoverLetter["Tailored Cover Letter (EN/HE)"]
        Sys2 --> Pitch["Recruiter Pitch"]
        Sys2 --> InterviewPrep["Anticipated Interview Questions"]
    end

    subgraph Autonomous Application Engine
        HighMatch --> Dispatcher["HybridApplicationDispatcher"]
        TailoredCV & CoverLetter --> Dispatcher
        Dispatcher --> Guardrails{"Safety Guardrails<br/>Score ≥ 85 / Israel Only / Daily Cap"}
        Guardrails --> StrategyRouter["Strategy Selector"]
        
        StrategyRouter --> EasyApply["EasyApplyStrategy<br/>Multi-Step Traversal (Up to 8 Steps)"]
        StrategyRouter --> BrowserStrategy["BrowserPlaywrightStrategy<br/>Comeet iFrame, Greenhouse, Lever, Workday"]
        StrategyRouter --> ApiStrategy["ApiPostStrategy<br/>Direct ATS JSON Dispatch"]

        EasyApply & BrowserStrategy & ApiStrategy --> Receipts{"4-Layer Verification Receipts"}
        Receipts --> L1["Layer 1: DOM Success Text"]
        Receipts --> L2["Layer 2: Redirect URL"]
        Receipts --> L3["Layer 3: Network HTTP 200/201"]
        Receipts --> L4["Layer 4: Timestamped Screenshot"]

        Receipts --> Ledger[("ApplicationLedger (SQLite)<br/>receipt_details JSON Proof")]
    end

    SeniorityCap --> Cache[Unified JobCache - 2h TTL]
    Cache --> Tools

🌟 Key Features

1. 11 Parallel Pluggable Sources:

   • Comprehensive multi-source aggregation across Comeet, Greenhouse, Lever, Workday, Eightfold.ai, DirectTech (Google, Apple, Amazon, IBM), LinkedIn, HireMeTech, AllJobs, GotFriends, and Jobify.

   • Resilient concurrency with semaphore throttling, rate limiting, and cross-platform deduplication.

2. System 1 Neural Triage Engine (<1ms Inference):

   • Local AI Youth Assistant (LAYA): Specialized fine-tuned model (based on Llama-3.2-3B) performing sub-millisecond candidate-job semantic alignment.

   • Seniority & Mismatch Capping: Immediately gates student/junior profiles from senior/lead positions, eliminating wasteful downstream processing.

   • Pluggable Architecture: Dependency-injected scoring engine supports laya (fine-tuned), base_laya (zero-shot baseline), generative (LLM gateway), or heuristic rule-based scoring.

3. System 2 Cognitive Tailoring:

   • Generates bespoke, job-specific application artifacts grounded with Pydantic models:

     • Tailored CV Highlights: Specific bullet points mapped to job requirements.

     • Custom Cover Letters: Professional, bilingual (English/Hebrew) cover letters matching company culture.

     • Recruiter Outreach Pitches: Compelling elevator pitches for recruiters and hiring managers.

     • Interview Preparation: Anticipated technical and behavioral questions based on candidate gaps and job specs.

4. Universal Auto-Apply & 4-Layer Verification Receipts:

   • Multi-Step LinkedIn Easy Apply: Navigates multi-step forms (up to 8 steps) with dynamic semantic form answering.

   • Universal Browser Solver: Handles nested iframe architectures (e.g. Comeet embedded widgets), modal overlays, and dynamic ATS flows.

   • 4-Layer Verification Receipts:

     • Layer 1 (DOM): Confirmation text matching ("Application submitted", "Thank you").

     • Layer 2 (URL): Redirect validation to confirmation paths (/thank-you, /submitted).

     • Layer 3 (Network): Intercepts HTTP 200/201 responses on submission endpoints.

     • Layer 4 (Screenshot Proof): Captures timestamped visual proof to disk (./data/receipts/*.png).

5. Dynamic Candidate Profiling:

   • Ingests .pdf, .docx, and .txt resumes for any technical discipline.

   • Derives technical skills, primary stack, and dynamic target roles without hardcoded templates.

6. Safety Guardrails & Application Ledger:

   • Immutable audit trail in SQLite (application_ledger.db) storing full submission receipts.

   • Strict fail-closed master switch (AUTO_APPLY_ENABLED=false).

   • Location enforcement (Israel/Remote only) and daily submission quotas (MAX_DAILY_APPLICATIONS).

🛠️ Tool Reference (16 Tools)

⚡ Quick Start

1. Clone & Configure

git clone https://github.com/zvieli/TechJobMCP.git
cd TechJobMCP

# Copy your CV and setup environment
cp /path/to/your/resume.pdf ./cv.pdf
cp .env.example .env

2. Run with Docker Compose

The Docker setup includes CPU-optimized PyTorch wheels and mounts local ./data for the fine-tuned LAYA model and persistent ledger:

docker compose up -d
docker compose logs -f techjob-mcp

2.1. Export Public HTTPS Port for AI Clients

curl -L --output cloudflared https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64
chmod +x cloudflared
./cloudflared tunnel --url http://localhost:8000

Connect the generated https://<tunnel-id>.trycloudflare.com/mcp URL to your AI client. See AI Client Integrations Guide for full setup instructions.

3. Dedicated CLI Tools

TechJobMCP provides purpose-built CLI scripts for autonomous scouting and evaluation:

# 1. Run complete end-to-end autonomous job hunt (Scout -> System 1 Score -> System 2 Tailor -> Auto-Apply)
uv run python scripts/run_autonomous_job_hunt.py --cv cv.pdf --limit 10

# 2. Scout real live job openings in Israel across all sources
uv run python scripts/scout_real_jobs.py --cv cv.pdf

# 3. Benchmark System 1 engines (Fine-Tuned LAYA vs. Base LAYA vs. Heuristic)
uv run python scripts/benchmark_system1_ab.py --sample-size 50

🧪 Running Tests

Run the full automated test suite (948 unit and integration tests):

uv run pytest

📄 License

Licensed under the MIT License. Copyright (c) 2026 Lior Zvieli.