JobWise

Job search with an agent-native interface. One engine, two surfaces: an MCP server any AI agent can drive, and — next — a web app.

JobWise reads open roles from public ATS APIs, scores them against your resume, and tracks your applications. The scoring explains itself, and nothing is scraped.

Status: Phase 0 complete — core engine and MCP server. Web app is next. See ROADMAP.md.

Why MCP

Job tools are built for a human clicking through a UI. That makes them invisible to the agents people increasingly work through.

JobWise exposes its engine over the Model Context Protocol, so you can ask your assistant "what should I apply to this week?" and it queries your real pipeline — no browser, no copy-paste, no separate app to remember to open.

Quickstart

git clone https://github.com/HadarAlfasi/jobwise.git
cd jobwise
npm ci
npm run build
npm run demo

npm run demo scans the bundled boards, filters and ranks them against a sample resume, and prints the top matches. No configuration, no API keys, nothing to sign up for.

Then wire it into your agent:

claude mcp add jobwise -s user -e JOBWISE_DIR="$PWD" -- node "$PWD/packages/mcp/build/index.js"

Restart your client and ask it what you should apply to.

What a session looks like

You:  What should I apply to this week?

      → search_roles { limit: 4 }

      42 matching role(s), showing 4:

      71  Payoneer — Team Lead R&D, CLM
          Ramat HaSharon, Tel Aviv District, Israel · posted 2026-08-19
          https://www.payoneer.com/careers/position/8146...
          · Title matches "Team Lead"
          · Seniority matches your target
          · Mentions 8 of your 31 skills: agile, go, kafka, microservices…

      70  Riskified — Software Engineering Manager
          Tel Aviv-Yafo, Gush Dan, Israel · posted 2026-08-11
          · Title matches "Software Engineer"
          · Seniority matches your target
          · Mentions 8 of your 31 skills: agile, aws, ci/cd, kubernetes…

You:  Anything I should be chasing?

      → needs_attention { silentForDays: 7 }

      2 application(s) quiet for 7+ days, longest first:

         30d  Northwind — Senior Backend Engineer   [applied]  id=app-northwind
         11d  Meridian Labs — Platform Team Lead    [applied]  id=app-meridian

You:  Meridian got back to me, I have a phone screen.

      → record_event { applicationId: "app-meridian", type: "interview",
                       note: "phone screen booked" }

      Recorded "interview" on Meridian Labs — Platform Team Lead.
      Status is now interview with 2 events on record.

(Roles above are real listings from the bundled boards; the applications are from the sample tracker in fixtures/.)

Tools

Tool	What it answers
search_roles	"What should I apply to?" — scans, filters, ranks, with reasons
pipeline_status	"Where does my search stand?" — counts by stage, what moved last
needs_attention	"What should I chase?" — applications gone quiet, terminal ones excluded
record_event	"Mark this as rejected" — append-only status change

Plus a jobwise://pipeline resource: every application with its full history.

record_event only writes to your local store. JobWise never submits an application, sends a message, or contacts an employer.

Configuration

Two files, both plain YAML.

config/boards.yml — the company boards to scan. One entry per board; adding a company is one line and needs no code change.

boards:
  - { provider: greenhouse, slug: catonetworks, company: Cato Networks }
  - { provider: ashby,      slug: finout,       company: Finout }

It ships seeded with Israeli tech employers verified to run a scannable board. Replace them with yours — find the slug in a company's careers URL (job-boards.greenhouse.io/SLUG, jobs.lever.co/SLUG, jobs.ashbyhq.com/SLUG).

config/profile.yml — your preferences. Copy the example:

cp config/profile.example.yml config/profile.yml

Set your location rules, target titles, and the path to your resume. config/profile.yml and my-resume.* are gitignored, so your real job search never lands in version control.

One trap the example warns about, because it cost real applications: location alwaysAllow is evaluated before block. Put a country there and every block rule silently stops applying — you get roles from cities you explicitly excluded, with nothing telling you why. Keep it to your home city. A regression test pins the behaviour.

Coverage, honestly

JobWise reads Greenhouse, Lever and Ashby through their public JSON APIs. No key, no scraping, no rate-limit games.

That is not everything. Many large employers — Comeet-hosted startups, and companies on Workday, iCIMS or a bespoke careers page — expose no scannable API, and no tool of this kind can see them. Support for more ATS platforms is on the roadmap; the ones already listed there are all plain HTTP.

The ATS platforms also publish no index of who uses them, so "scan everything" is a discovery problem rather than a fetching one. See the roadmap for the approach.

No scraping, by design

JobWise does not scrape job boards, and will not. It is against their terms, it breaks constantly, and on a tool aimed at recruiters it is the wrong signal. Fetching a single public posting a user explicitly pastes is a different act and is planned; harvesting a board is not.

How it works

  Greenhouse ┐
  Lever      ├──▶  @jobwise/core  ──┬──▶  MCP server   (agents drive it)
  Ashby      ┘   fetch · normalize  └──▶  web app      (next)
  your resume ─▶ match · track

All the logic lives in @jobwise/core. The surfaces are thin adapters that translate and format — if behaviour appears in one of them, it belongs in the core.

Scoring is deterministic and explainable. Five weighted components — title, skill overlap, seniority, recency, home location — each contributing points with a human-readable reason. No model call, no API key, same inputs always give the same score. That is deliberate: it is the baseline any future LLM-based scoring has to beat measurably, rather than being assumed to be better.

A component with no data abstains rather than scoring zero, so a board that publishes no dates does not quietly rank below a chattier one.

Application history is append-only. Status is derived from the event log rather than stored, so history can never disagree with current state. A correction is another event, never an edit.

Development

npm run build      # compile both packages
npm run typecheck  # includes test files
npm test           # 204 tests, no network
npm run demo       # end-to-end against live boards

The suite blocks fetch entirely (test/no-network.ts), so an ATS outage can never turn the build red. Provider tests run against committed fixtures captured from real responses — refresh them with node scripts/capture-fixtures.mjs.

stdout is the JSON-RPC channel. A console.log anywhere in core or mcp corrupts the MCP stream, and the web app would tolerate the same line happily. A test scans the source and fails the build on it, and the integration test spawns the real server and asserts every line it writes to stdout is valid JSON-RPC.

Credits

The tracker importer reads the file format of career-ops by Santiago Fernández de Valderrama, so its users can seed JobWise with existing history in one command. JobWise is separate work and shares no code with it.

License

MIT © Hadar Alfasi