Job Application MCP

A local-first, open-source job application assistant built on the Model Context Protocol.

Analyse jobs · Match against your CV · Tailor documents · Preview forms · Track applications — all on your machine, nothing uploaded, submission stays in your hands.

Table of contents

• Why

• What it does (free core)

• What it deliberately does NOT do

• How it works

• Quick start

• Connect an MCP client

• Tools reference

• A full example flow

• Configuration

• Project layout

• Companion apps

• Develop

• How matching works

• Privacy & data

• Roadmap / Pro upgrade

• Limitations

• License

• Contributing

Related MCP server: jobjourney-claude-plugin

Why

Job hunting is repetitive: read a posting, check it against your CV, tweak a cover letter, fill the same fields on yet another form, and remember where you applied.

Job Application MCP turns that into a set of tools any MCP-compatible AI client (Claude Desktop, Claude Code, others) can drive for you — while keeping you in control:

• Your CV and applications never leave your machine in the free core.

• The assistant prepares and previews; you review and submit.

• It's open source under AGPL-3.0, so you can audit exactly what it does.

ℹ️ This repository is the free community core described in BUSINESS_PROPOSAL.md. Paid AI/cloud/business services are intentionally out of scope here.

What it does (free core)

What it deliberately does NOT do

• ❌ No automated submission. You submit manually on the employer's site.

• ❌ No automation of LinkedIn or Indeed — their Terms of Service prohibit it. Paste the description and complete those by hand.

• ❌ No CAPTCHA solving, 2FA bypass, or login recovery — it pauses for manual action.

• ❌ No AI rewriting, cloud sync, or premium adapters — those are future Pro/cloud services.

These aren't missing features; they're the design.

How it works

┌─────────────────┐      stdio (JSON-RPC)      ┌──────────────────────┐
│  MCP client     │  ◀────────────────────────▶│  job-application-mcp │
│  (Claude, etc.) │                             │  (this server)       │
└─────────────────┘                             └──────────┬───────────┘
                                                           │ reads / writes
                                                           ▼
                                                ┌──────────────────────┐
                                                │  local SQLite DB     │
                                                │  + parsed CV text    │
                                                │  (JOB_MCP_DATA_DIR)  │
                                                └──────────────────────┘

You (or an AI client) call tools. Each tool does a small, auditable thing and returns structured JSON. Tools that touch external sites do not exist; everything operates on data you provide and store locally.

Quick start

Requirements

• Node.js ≥ 22 (uses the built-in node:sqlite — no native compilation needed)

• npm

Install & build

git clone https://github.com/<your-user>/job-application-mcp.git
cd job-application-mcp
npm install
npm run build      # -> dist/

Verify it works

npm test           # full unit/integration test suite
node dist/cli/cli.js --help

Connect an MCP client

Build the project, then add the server to your MCP client's config.

Claude Desktop / Claude Code

{
  "mcpServers": {
    "job-application-mcp": {
      "command": "node",
      "args": ["/ABSOLUTE/path/to/job-application-mcp/dist/src/index.js"]
    }
  }
}

⚠️ Use an absolute path — MCP clients launch the server from their own working directory, not yours.

Run directly (no client)

node dist/cli/cli.js serve      # equivalent to: node dist/src/index.js

The server speaks JSON-RPC over stdio. See A full example flow for raw request examples.

Tools reference

All 23 tools are local and synchronous-ish (CV parsing and AI calls are async). Arguments are validated with zod; the MCP client receives a JSON Schema for each. The 13 free-core tools (profile → autofill) are the everyday workflow. The credits/Pro and admin/business tools are wire-ready seams for the hosted Pro service (see Business model); they run locally today and do not require any subscription.

Profile

CVs

parse_cv supports .pdf (pdf-parse), .docx (mammoth), and plain text. If you pass text, it's stored as-is. file_path must be inside the data dir (or a folder listed in JOB_MCP_CV_DIRS) — paths outside the allow-list are rejected, so an MCP client can't read arbitrary files.

Jobs

Matching & prep

AI tools and your own key. tailor_cv, cover_letter, and draft_answer use a real AI provider when you set AI_PROVIDER (openai or anthropic) and AI_API_KEY — on the Free or Pro plan. Without a key they fall back to a local heuristic. The Pro hosted path additionally debits an AI credit; using your own key is never debited. See Business model.

Applications & forms

Status pipeline: draft → ready → submitted → interview → offer → rejected → closed.

Analytics & status

Credits & Pro (wire-ready seams)

Admin / business (wire-ready seams)

A full example flow

Here's a complete session using the stdio JSON-RPC interface directly. (An MCP client like Claude does this for you in natural language.)

node dist/src/index.js

Send these lines (one JSON-RPC object per line):

// 1. handshake
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"demo","version":"0"}}}
{"jsonrpc":"2.0","method":"notifications/initialized"}

// 2. set up your profile
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"update_profile","arguments":{"full_name":"Ada Lovelace","headline":"Backend Engineer","skills":["python","sql","aws","docker"]}}}

// 3. analyse + store a job (paste the description text)
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"analyze_job","arguments":{"title":"Backend Engineer","company":"Acme Co","description":"We need a backend engineer with Python, SQL, and AWS to build data pipelines. PostgreSQL and Docker are a plus."}}}

// 4. parse a CV (file path or pasted text)
{"jsonrpc":"2.0","id":4,"method":"tools/call","params":{"name":"parse_cv","arguments":{"label":"main-cv","text":"Ada Lovelace — Backend Engineer. Python, SQL, AWS, PostgreSQL. Built ETL pipelines..."}}}

// 5. score the CV against the job (save=true creates a draft application)
{"jsonrpc":"2.0","id":5,"method":"tools/call","params":{"name":"match_cv","arguments":{"job_id":1,"cv_id":1,"save":true}}}

// 6. preview autofill for a form (no submission)
{"jsonrpc":"2.0","id":6,"method":"tools/call","params":{"name":"autofill_form","arguments":{"application_id":1,"form_fields":[{"name":"name","label":"Full name"},{"name":"email","label":"Email"},{"name":"salary","label":"Salary expectation"}]}}}

// 7. after you submit on the employer's site, mark it submitted
{"jsonrpc":"2.0","id":7,"method":"tools/call","params":{"name":"update_application_status","arguments":{"application_id":1,"status":"submitted"}}}

Step 6 returns a mapping where the salary field is flagged requires_user_review: true — sensitive fields always need your eyes.

Configuration

All config is via environment variables. The free core needs none of them.

The following enable your own AI key (works on Free or Pro) and are placeholders for future hosted Pro/cloud services (see .env.example):

Project layout

job-application-mcp/
├── src/
│   ├── index.ts           # stdio MCP server entrypoint
│   ├── http.ts            # local HTTP bridge entrypoint (127.0.0.1)
│   ├── server.ts          # server factory + tool registry
│   ├── tools/             # task and compatibility MCP tools
│   │   ├── types.ts       # ToolDef / AnyTool / result helpers
│   │   ├── profile.ts     # get/update_profile
│   │   ├── cv.ts          # parse_cv, list_cvs
│   │   ├── job.ts         # analyze_job, get_job
│   │   ├── matching.ts    # match_cv, tailor_cv, cover_letter, draft_answer
│   │   ├── application.ts # save/list/update_application, autofill_form
│   │   ├── analytics.ts   # application_analytics, status
│   │   ├── pro.ts         # credits, topup_credits, grant_monthly_credits
│   │   └── admin.ts       # admin_create_account / add_candidate / list_team / usage_report
│   ├── store/             # SQLite layer (node:sqlite)
│   │   ├── db.ts          # open / migrate / close / reset
│   │   ├── profile.ts     # profile CRUD
│   │   └── applications.ts# cv / job / application CRUD
│   ├── cv/parser.ts       # PDF / DOCX / TXT parsing (path allow-listed)
│   ├── ai/                # provider abstraction (mock / openai / anthropic)
│   ├── licence/           # entitlement tokens + credits + referrals
│   ├── payments/          # payment-webhook seam (server-side, not in clients)
│   ├── sync/              # cloud-sync seam (local no-op until Pro)
│   ├── features.ts        # feature gating (plan + credits + AI mode)
│   └── lib/
│       ├── types.ts       # domain types
│       ├── crypto.ts      # HMAC-SHA256 + constant-time compare
│       ├── entitlement.ts # Entitlement value type + plan limits
│       └── scoring.ts     # keyword extraction + match scoring (pure)
├── cli/cli.ts             # human CLI: job-mcp serve | serve:http | --help
├── extension/             # Chrome MV3 extension (form capture → preview)
├── desktop/               # Electron wrapper UI (launches the bridge)
├── .github/workflows/ci.yml  # build + typecheck + test + stdio smoke
├── .github/workflows/daily-effectiveness.yml # scheduled regression gate
├── evals/                 # synthetic effectiveness fixtures + thresholds
├── tests/                 # node:test suite
├── BUSINESS_PROPOSAL.md   # open-core + paid-service model
├── CLAUDE.md              # guide for Claude Code working in this repo
└── LICENSE                # AGPL-3.0-or-later

Companion apps

The free core is the stdio MCP server above. Two optional companions live in this repo and talk to a local HTTP bridge (same tools, loopback only):

HTTP bridge

npm run serve:http      # 127.0.0.1:8787 by default

Endpoints: GET /health, POST /call with {"name","arguments"}. Optional bearer auth via JOB_MCP_HTTP_TOKEN. Source: src/http.ts.

Chrome extension — extension/

Captures form fields on a career page and asks the bridge to preview an autofill mapping. Never submits. Load it unpacked from extension/. See extension/README.md.

Desktop app — desktop/

An Electron wrapper that launches the bridge and shows a dashboard (status, profile, applications). Run with cd desktop && npm install && npm start. See desktop/README.md.

Both companions are read/preview only and reach only 127.0.0.1. They inherit the project's no-submission, no-LinkedIn/Indeed-automation rules.

Develop

npm run build       # tsc -> dist/
npm run typecheck   # tsc --noEmit
npm test            # node --test (scoring + store)
npm run eval:effectiveness # synthetic safety/effectiveness report
npm run dev         # tsc --watch

GitHub also runs the deterministic effectiveness gate daily at 01:17 UTC (09:17 Malaysia time). It uses synthetic fixtures, makes no model calls, and never changes code or user data automatically. See docs/EFFECTIVENESS.md for metrics and limits.

Adding a tool

1. Create src/tools/<name>.ts exporting a ToolDef<z.ZodObject<…>>.

2. Register it in src/tools/index.ts (add to the tools array).

3. Add a test in tests/ if it has pure logic worth covering.

Tools receive parsed, typed input and return a ToolResult ({ summary, data?, notes? }). The server handles JSON Schema generation, validation, and error wrapping for you.

How matching works

Matching is a pure, heuristic function in src/lib/scoring.ts — no AI, no network.

1. Extract keywords from the job description: tokenize, drop stopwords and filler verbs, weight tech hints and frequent terms.

2. Score: for each job keyword, check the candidate's curated profile skills plus skills detected in the CV text. Stem-ish matches count (e.g. react ≈ reactjs).

3. Result: score (0–100), matched, missing, extra, and notes (e.g. "Below 50% — consider tailoring or skipping").

Formula: score = matched/total × 100 plus a small, capped breadth bonus for extra skills (only when there are job keywords). It's intentionally simple and auditable — a baseline you can replace with the Pro AI scorer later.

Privacy & data

• Local-first by default. CVs, profile, and applications live in JOB_MCP_DATA_DIR (default ./data), which is gitignored. Ollama and the heuristic stay local; configuring OpenAI, Anthropic, an update feed, or a licence endpoint enables the corresponding outbound request.

• Provider keys are not stored in SQLite. Desktop-entered keys last only for the bridge process; environment variables can provide restart persistence.

• No PII in logs. CV text and personal fields aren't logged at info level.

• Sensitive fields are flagged. autofill_form marks salary, authorization, gender, disability, and consent fields as requires_user_review: true.

Windows alpha installer

GitHub prereleases provide a standalone x64 one-click Setup executable. The installer bundles the local bridge, so a separate Node.js installation is not required. Version 0.4.2 is intentionally unsigned alpha software: Windows SmartScreen may display a warning. Verify the downloaded file against SHA256SUMS.txt from the same GitHub prerelease before running it. This is not a production-ready hosted or paid release.

• You can wipe everything by deleting the data directory.

Roadmap / Pro upgrade

The free core is the foundation. BUSINESS_PROPOSAL.md describes the open-core model. Likely next steps:

• Free core: match_cv + tailor_cv integration test through the MCP layer; richer keyword extraction (skill taxonomy).

• Distribution: Windows installer via GitHub Releases; Chrome Web Store extension for form-field capture.

• Pro (separate hosted services): AI CV tailoring & cover letters, encrypted cloud backup, multi-device sync, maintained site adapters, application analytics.

• Business: multi-candidate accounts, team dashboards, white-label.

Tool handlers are written so a Pro upgrade can slot in without changing the public tool surface — e.g. tailor_cv returns heuristic suggestions today and could call an AI service tomorrow.

Limitations

• Keyword extraction is heuristic and will miss or mis-rank some skills. Confirm against the source posting before relying on a match score.

• No browser automation. Form fields must be supplied to autofill_form (manually or, later, via the extension). The tool previews; it doesn't fill.

• LinkedIn / Indeed: analysis-only, paste the description, submit manually.

• Alpha status: APIs and stored schemas may change before 1.0.

License

Copyright © 2026 Job Application MCP contributors.

This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. See LICENSE.

A separate commercial licence is available for embedding without AGPL obligations — see BUSINESS_PROPOSAL.md §8.

Contributing

Issues and PRs welcome. By contributing you agree your contributions will be licensed under the same AGPL-3.0-or-later terms.

Please keep the free core fully local and free — no feature in this repo may gate a free-core capability behind a paid/network service. See CLAUDE.md for the conventions used in this codebase.