job-search

▶ Try it now: https://job-search-mcp-tau.vercel.app No sign-up, no keys. Fill in a profile, search jobs, generate a cover letter, and rehearse a Q&A.

Where I pull the jobs from

When you tick "Include live listings", I fetch from five keyless sources in parallel, filter them by your profile's role and location, then merge, de-duplicate, rank, and check that every link is actually reachable before I show it to you.

If you leave live listings off, search runs instantly over a bundled, illustrative sample dataset.

Related MCP server: mcpforwork

The profile you give me

Everything keys off a simple candidate profile. I keep it only in your browser (localStorage) and pass it inline to each call, so the server stays stateless — your data never sits on my backend.

What you can do — demo vs. live

The live deployment runs in Live AI mode through Groq (llama-3.3-70b-versatile), so letters and answers are model-generated. The banner in the UI tells you whether you're in Demo or Live AI mode.

Performance — how I keep live search fast

Hitting five APIs and validating ~20 links on every request is slow (~3–4s) and abuses the sources. So I cache two things: the merged source results (keyed by role + location, 10-minute TTL) and each link's reachability (validated once, then reused). A repeat search does zero outbound calls and returns the same results almost instantly.

Measured locally, same query, same results:

That's roughly a 45× speed-up on the warm path — and I keep correctness, so I never serve a link I haven't verified. To stop users ever paying the cold cost, there's an off-request-path warm-up endpoint, /api/cron/revalidate, you can put on a schedule (Vercel Cron or any pinger). The cache is in-process by default (zero config); set a Vercel KV / Upstash store (KV_REST_API_URL + KV_REST_API_TOKEN) to share it across every instance.

Use it as an MCP server

This app is a remote MCP server, so you can connect your MCP client (Claude Desktop, Claude Code, Cursor, …) and let the model fetch jobs for a candidate.

• Endpoint: https://job-search-mcp-tau.vercel.app/api/mcp (Streamable HTTP + SSE)

• Tools: profile_upsert, jobs_search, letter_generate, qa_reply

{
  "mcpServers": {
    "job-search": { "url": "https://job-search-mcp-tau.vercel.app/api/mcp" }
  }
}

Ask your assistant: "Find remote data-engineering roles for someone strong in Python, Spark and SQL" → it calls jobs_search and gets back ranked, link-checked jobs with fit scores.

REST API

curl -s -X POST https://job-search-mcp-tau.vercel.app/api/jobs \
  -H "Content-Type: application/json" \
  -d '{"query":"python data engineer","profile":{"skills":["python","spark","sql"]},"live":true,"limit":5}'

Run it locally

git clone https://github.com/VikramKavuri/Jobsearch_using_MCP_server.git
cd Jobsearch_using_MCP_server
npm install
npm run dev      # http://localhost:3000
npm test         # 77 unit tests (pure functions, no network)

You don't need a .env — it starts in demo mode. To turn on live AI, copy .env.example to .env.local and set one key (GROQ_API_KEY, ANTHROPIC_API_KEY, OPENAI_API_KEY, or HF_TOKEN).

Deploy your own

npm i -g vercel
vercel --prod    # prompts for login the first time

Vercel auto-detects Next.js. Add an API key under Project → Settings → Environment Variables to enable live AI, then redeploy.

How I built it

app/
  page.tsx                       Web UI: 4 tabs (Profile, Job Search, Cover Letter, Q&A)
  api/{config,profile,jobs,letter,qa}/route.ts   thin REST adapters
  api/[transport]/route.ts       MCP endpoint (4 tools) at /api/mcp
  api/cron/revalidate/route.ts   off-path cache warming
lib/
  tools/{profile,search,letter,qa}.ts   pure capability functions (+ unit tests)
  ranking.ts                     TF-IDF cosine over job text (pure TS)
  jobs-source.ts                 5 live sources + bundled sample, mappers, dedupe
  link-check.ts                  reachability validation for live job links
  cache.ts                       in-memory + Vercel KV cache, one tiny interface
  config.ts                      env → real-vs-demo decision (the only env reader)
  llm.ts                         provider abstraction (Groq / OpenAI / Anthropic / HF ↔ demo)
  service.ts                     composition root shared by REST + MCP

I kept the capability functions in lib/tools/* and lib/ranking.ts pure — no Next, no env, no network — so I can unit-test them in isolation. lib/config.ts is the only place that reads env and decides demo-vs-live; the tools receive an injected llm and never branch on the environment. REST and MCP both call lib/service.ts, so the two surfaces can't drift.

Deeper dive: docs/ARCHITECTURE.md has the data flow, my design trade-offs, and an honest look at what I'd change to push this further.

Engineering highlights

• One core, three surfaces. I made the web UI, REST, and MCP thin adapters over a single composition root (lib/service.ts) — zero duplicated logic, so the surfaces can't drift.

• Testable by construction. Ranking and the four capabilities are pure functions; 77 deterministic unit tests run offline (Vitest) and on CI on every push.

• Fast where it matters. Two-layer caching (source results + link reachability) takes a repeat live search from ~3.7s to ~80ms while keeping every link verified.

• Resilient by design. I fetch five sources in parallel and each degrades to [] on failure; results are de-duped, ranked, and link-checked — one dead source or dead link never breaks search.

• Pluggable AI. A provider abstraction (lib/llm.ts) swaps Groq / OpenAI / Anthropic / HF behind one interface, with a deterministic demo path so nothing requires a key.

Notes

• Attribution: live job data comes from Remotive, The Muse, Arbeitnow, RemoteOK and Jobicy. RemoteOK and The Muse ask that you credit them when you display their results — so I do.

• Stateless by design — there's no database; your profile lives in your browser.

• This is my clean Vercel rebuild of the original Hugging Face Spaces "Job Search MCP" concept (no torch / faiss / Gradio).

License

MIT © VikramKavuri