求职追踪器

Job360

The job tracker your AI assistant fills in for you — every CV version, every reply, every receipt.

Your AI assistant (Claude, ChatGPT, Claude Code, any MCP client) does the work of a job hunt: it finds the job, judges the fit, writes the CV and cover letter, reads your Gmail, fills the form. Job360 is where all of that is recorded — the profile it works from, every document version it wrote, every event with who wrote it and when, and a receipt you can trust when you apply.

Job360 has no AI of its own. It never searches for, ranks or recommends jobs (product rule 4), and it never judges you — it stores what your assistant decided and shows it back to you. Read docs/product/VISION.md first.

Live at https://job360.uk (Railway, auto-deployed from main).

How it works

 You ──► your AI assistant ──(MCP)──► Job360 ──► the record you see on job360.uk
         (thinks + acts)               (stores)

1. Connect your assistant once — the address is https://job360.uk/api/mcp. Connector apps (Claude.ai is proven end to end; ChatGPT-style connectors use the same flow) sign in with OAuth 2.1; CLI clients like Claude Code use a personal token (j360_…, shown once, stored hashed).
2. Type run 360. The assistant fetches the setup recipe from Job360, reads your CV, fills your profile, asks to connect Gmail, and offers the inbox check.
3. Hunt and apply. The assistant searches with its own tools, brings each job it judges a fit, writes a tailored CV + cover letter, and stops before Submit. You say yes; it records the application and Job360 freezes a receipt.
4. Follow up. On its schedule the assistant reads Gmail, records replies, interviews and rejections against the right application, sends the outreach emails it wrote (Auto mode only), and tells you what is due and what has gone quiet.
5. Look back any time. Ask your assistant "how is my hunt going?" or "tell me everything about the Mistral application" — it reads it all from Job360. Or open the web app.

What is live today

For your assistant (MCP server at /api/mcp)

• Recipes — run 360 and friends. Step-by-step playbooks the assistant fetches with get_recipe; which ones exist, and the order a new user runs them in, is api/routes/recipes.RECIPE_NAMES (text in backend/src/recipes/).
• Profile — get_profile returns the raw CV / LinkedIn / GitHub text plus the structured fields; the assistant writes skills, dated work history, projects, targets and preferences back with update_profile. Every agent edit is kept with its author and can be taken back.
• Bring a job — bring_job stores the ad (link or pasted text) and births one Application, with the job's country, remote flag and where it was found.
• Fit and visa — the assistant's own verdict, stored with save_fit (Job360 never computes a score).
• Documents — save_artifact keeps every CV, cover-letter, answer and outreach version, with the assistant's ATS score and notes on each CV.
• Events — record_event appends a typed history: replied, interview requested / scheduled / done, offer, rejected, follow-ups with dates, notes — each with its source (e.g. the Gmail message id) so a re-read never double-counts.
• Receipts — record_application freezes exactly what was sent (CV version, cover letter, answers, channel, confirmation) the moment you say you applied. Never edited afterwards.
• People — recruiters and hiring managers with where they were found, every outreach message version, sent and replied marks.
• Needs you — when the assistant would have to guess, it calls ask_user; you answer once, on the web or in chat, and every assistant reads the answer.
• Inbox check — your choice, stored in Job360 so every assistant obeys it: Auto (reads Gmail for your open applications, records what happened, sends the outreach emails it wrote) · Ask me first (asks before each check, only drafts) · Off. How often: the choices are services/profile/models.VALID_CHECK_EVERY_VALUES.
• Read-back — get_application (one job, everything), export_history (the whole hunt), stats, list_people, list_receipts, whats_new.

For you (web app at job360.uk)

Every page and the one question it answers: docs/product/VISION.md → "Every page has one purpose". Light and dark mode; magic-link or password login.

What Job360 deliberately does not do

• Search, rank, score or recommend jobs.
• Run its own LLM, or read your email itself — your assistant does both with its own connectors.
• Submit an application without your yes for that one application, or follow instructions written inside an email.
• Push notifications or run background jobs — it is pull-based; your assistant's own scheduled task does the checking.

Architecture

Two deployables on one Postgres database:

• backend/ — FastAPI (Python 3.12 in prod). Product path: POST /api/jobs/bring (api/routes/bring.py) → the application spine (services/applications/spine.py: one Application, append-only events, versioned artifacts, frozen receipts) → the MCP server (api/mcp_server.py), with services/profile/ feeding every step. Migrations apply automatically on boot.
• frontend/ — Next.js 16 + React 19 + Tailwind 4, a thin screen over the same routes.

The deep reference (directory tree, schema, dependencies) is ARCHITECTURE.md. Counts (routes, migrations, workflows) are never written down: measure it (ls, grep -c, python -m pytest --collect-only -q). Interactive API docs: http://localhost:8000/docs when the backend runs.

Quick start (local)

git clone https://github.com/Ranjith36963/job360.git
cd job360

# Postgres (host port 5433)
docker compose -f docker-compose.dev.yml up -d postgres

# Config
cp .env.example .env          # then fill it in

# Backend — FastAPI on :8000, MCP at /api/mcp
cd backend
pip install -e ".[dev]"
python main.py

# Frontend — Next.js on :3000
cd ../frontend
npm install
npm run dev

Optional: bootstrap a single dev profile from the command line — python -m src.cli setup-profile --cv cv.pdf --linkedin linkedin.pdf --github yourname (all flags optional). Signed-in users build their profile on the web or through their assistant instead.

Testing

cd backend && python -m pytest -q -p no:randomly      # needs the dev Postgres; runs offline
cd frontend && npm run test:unit && npm run test:e2e

The backend suite runs on a real Postgres, schema per test, with HTTP mocked by aioresponses. Measure counts, never quote them: python -m pytest --collect-only -q -p no:randomly | tail -1. Before a commit, bash scripts/agent-gate.sh runs the targeted tests, lint and type checks for what you changed.

Deployment

main is production: every merge auto-deploys the backend, frontend and Postgres services on Railway. System email (magic links, password reset) goes through Resend on job360.uk.

Contributing

Branch, commit and PR conventions: CONTRIBUTING.md. Product rules: docs/product/product_design_rules.md. Current phase: STATUS.md.