Jobdork – job leads from employers' career sites, ranked by your resume

jobdork finds job posts for you, on your own computer. It reads employers' own hiring systems and job boards, keeps only the posts that pass the rules you set (job titles, place, pay, dealbreakers), ranks them against your resume, and keeps track of what you did with each one. An AI model can read the best ones against your resume and draft a cover letter or a tailored resume; every claim it writes is checked against your resume and the advert.

It is not a job board and has no account: nobody posts to it, and your resume and searches stay on your machine.

Start with the download for macOS or Windows, or another way to run it: from source, in a terminal, in Docker or on a home server. The user guide then walks through every page of the app with pictures. CONFIG.md has every setting, PRIVACY.md what jobdork keeps and every address it contacts, and SECURITY.md how releases are checked and how to verify a download.

For how it works underneath: where its coverage comes from (SOURCES.md), what each source answers and where it falls short (PLATFORMS.md), what a scan reaches region by region (REGIONS.md), how the pipeline and the AI reader fit together (ARCHITECTURE.md), the choices behind it (DECISIONS.md) and how it got here (HISTORY.md). Every change, and what is still outstanding, is in the changelog.

• Reads where jobs are posted. 420 employer boards ship built in (338 of them US employers, across 27 industries) on Workday, Oracle Recruiting Cloud, Eightfold, Taleo Business Edition, Greenhouse, Lever, Ashby, SmartRecruiters, Breezy and employers' own careers sites. Keyword sources need no account: Workable's search, Himalayas (remote work), Kalibrr (Philippines, Indonesia), Germany's employment agency and Arbeitnow. Adzuna and USAJOBS work with a free key of your own, andjobdork discover adds any other employer by reading the board off their careers page.Where the jobs come from
• Screens with your rules, and says why. Every post dropped is counted with its reason. Only pay the employer states can hide a post: one with no pay is kept and marked, since most posts state none.The rules
• Cleans up the list. Copies of one job are shown as one post; each post is aged from the board's own dates, with old ones ranked lower and ghost listings (over 90 days) dropped; an Adzuna post the employer's own board does not carry is flagged.
• Ranks against your resume , offline:Match (title, place, pay) andFit (the advert's skills you have and lack).Scores
• Checks your resume , offline: how an applicant tracking system reads the file and what to fix, how its lines are written, and each post'sATS match (keywords, required skills, title, sections, dates, out of 100).The Resume page
• Uses an AI model only when you set one up. Ollama on your own computer, or Anthropic, Gemini, OpenAI or Groq with your key, held in memory or, if you choose, in your computer's keychain; never in a file. Or your own coding agent (Claude Code, Codex, Antigravity CLI, Cursor, GitHub Copilot, Claude Desktop) does the writing through jobdork's MCP server or command line, with no key, within guardrails: jobdork checks and keeps what it writes, and what it adds to your record waits for your OK; the AI page and the Dashboard say when it last used jobdork's tools and what it did (Coding agent ). It can judge posts (Job posts marks which ones each run judged, and sorts and filters by its score), write a cover letter, review your resume, and write a tailored resume as a PDF and a Word file (in a sans or serif face), held to Indeed's length advice (one page up to ten years), interview prep (fifteen likely questions, with STAR outlines) and a recruiter message. Eight more tools help you prepare: a tailoring report, an interview story, a thank-you email, offer negotiation, an outreach message, a LinkedIn headline and About, changing field, and a 30-day search plan. A second model checks each output, claim by claim, against your whole resume and the advert, and anything unsupported is flagged: 2.9% of claims with the recommended local model, measured on seven resumes in six fields (models compared ). A flagged claim is then taken out or rewritten, and checked again, before you see it. It never applies or sends for you.The AI page
• Keeps track. Statuses fromnew tooffer , notes, a check for posts that have closed, scans on a schedule, notifications (a click opens jobdork), and an email digest.My applications keeps every job you applied to, by stage, with each status dated, the next step and its date, the contact, the resume and letter you sent (uploaded, and read by that job's AI tools), and notes. OneAdd a job post form adds a post from its link, already applied or not.My applications
• Reaches past what it can read. ASearch Google Jobs link opens your search in Google's job results, andjobdork dork builds Google searches for sites that cannot be read automatically (LinkedIn, Indeed and others).docs/DORK.md
• Works outside the US. A place list of 70,026 places in 245 countries, any country you name, and distances in miles or kilometres to match. What reaches each region is measured indocs/REGIONS.md .

What it is not good at: it takes some setting up; it reaches fewer employers outside North America (docs/REGIONS.md); it does not work around sites that refuse automated readers; and an advert that arrives cut short (Adzuna sends 500 characters) cannot be screened fully. Limits has the list.

From source, in three lines:

git clone https://github.com/jessn-dev/jobDork.git && cd jobDork
python3 -m venv .venv && .venv/bin/pip install -e '.[pdf]'
.venv/bin/python -m jobdork serve      # opens setup in your browser

The first start opens setup, one question per screen: your resume (your master resume, if you keep one), an AI model (optional), the job titles to look for, where you are, pay and job sources. It ends with a first scan. Getting started

Your settings live in config.yaml, which is not in the repository (it is gitignored, with .env, data/ and out/), so your search stays yours.

The installers are not code-signed, so macOS and Windows warn the first time you open one; the desktop guide says how to get past it and how to check a download instead, and gives a one-command install for the Mac that checks the download itself and skips the warning. Every installer carries a build attestation from this repository's release workflow, and the app's Update button opens a new installer only after that attestation verifies (SECURITY.md).

• jobdork sends nothing to its makers. It goes on the network only to do what you set up (read job sources, use the AI provider you chose, send feedback when you press Send) and, once a day unless you turn it off, to check for a newer release. Every destination:PRIVACY.md .

• The dashboard is yours alone. It listens on this computer only (in Docker, on the addresses you name and no others) and needs a token made fresh for each run; a job post's text and links are treated as hostile (what a job post cannot do ).

• Releases are checked before they are published: tests, vulnerability and secret scans on every change, dependencies pinned by hash, the Docker image signed (cosign) and the installers attested, and each release approved by hand.SECURITY.md

• Found a problem? Report it privately:SECURITY.md .

• Employers on systems with no reader (UKG, iCIMS, Jobvite, BambooHR, ADP and others) are reached only when their own careers site marks its posts up for search engines, or throughjobdork dork .docs/PLATFORMS.md

• Sites with no public way in, or that refuse automated readers (LinkedIn, Indeed, Glassdoor, SEEK) are not worked around: the Google Jobs link andjobdork dork cover them.

• Salary: only a minority of posts state pay, so a pay floor filters that minority and the rest are kept, markedunconfirmed salary .

• Right to work: a post that states its sponsorship position is flagged; most state nothing, so treat an unflagged post as unknown.

• An empty answer from a board is not proof it is hiring nobody: some answer the same way when throttling, and jobdork reports those as unknown.

• Jobs never posted online in a system it can read, common for small local employers, cannot be found.

Why was a post dropped? jobdork list --flags, then how a title is matched and the salary rule. Why did a source return nothing? Failure usually looks like success.

Contributions are welcome: CONTRIBUTING.md, and the code of conduct for everyone taking part. Issues start from a form: a problem, an idea or a question.

pip install -e '.[dev]'     # pytest, ruff, and sigstore for the update tests
pytest                      # the whole suite; none of it goes online
ruff check .                # lint
python tests/run_all.py     # the same tests, without pytest

python packaging/build.py builds the desktop installer for the machine it runs on (how). Dependencies are pinned with their hashes in requirements.lock (the Docker image) and requirements-desktop.lock (the installers).

.
├── jobdork/
│   ├── cli.py             # every command
│   ├── core/              # config, setup, schedule, notifications, updates, telemetry
│   ├── db/                # SQLite store, migrations, grouping copies of a job
│   ├── fetch/             # one reader per source, robots.txt, the paced HTTP client
│   ├── search/            # scan, screening, built-in boards, discover, freshness, places, resume, ATS match
│   ├── ai/                # the model, judging, the hallucination check, letters and reviews
│   ├── writing/           # drafts by Claude Code, tailored resumes, the checks every draft gets
│   ├── web/               # the dashboard server, its API, the desktop window, and a coding agent's tools (MCP)
│   ├── output/            # the static page and the email digest
│   ├── dork/              # the Google search builder
│   └── data/              # built-in boards, the place list, the dashboard page, humanizer rules
├── guide/                 # the user guide (MkDocs)
├── docs/                  # the documentation above
├── packaging/             # the desktop build: PyInstaller spec, icon, Windows installer
├── scripts/               # board list, place list, guide pictures, release notes, image check
├── tests/
├── Dockerfile, compose.yaml
└── .github/workflows/     # checks and releases (docker.yml, desktop.yml), the user guide, the weekly board re-check

docs/ARCHITECTURE.md explains each part; guide/README.md how the user guide is built and its pictures taken.

• humanizer by blader (MIT): the rules for spotting signs of AI writing, drawn from Wikipedia's "Signs of AI writing" guide. ItsSKILL.md is copied intojobdork/data/humanizer/ with its licence; drafting models are given it, andjobdork/writing/humanize.py checks drafts for the patterns a script can find.
• HalluLens (Bang et al.,arXiv:2504.17550 ): the method the hallucination check follows, splitting an output into single claims and checking each against its source. Only the method is used; none of its code, prompts or data (CC BY-NC) is included.
• Resume Matcher by srbhr (Apache-2.0): the resume readability check (jobdork/search/readable.py ), the writing rules (jobdork/writing/style.py ) and the ATS match (jobdork/search/ats.py ) follow its parse check, writing check and ATS score, their rules, thresholds and weights, and the interview prep and recruiter message its prompts' parts, rewritten for jobdork; none of its code is copied. The Resume page's layout follows its dashboard.
• Place names from GeoNames (CC BY 4.0); icons fromLucide (ISC).

GNU General Public License v3.0 or later: you may use, change and share jobdork, and anything you distribute that is built from it must be shared under the same licence, with its source. The humanizer rules keep their MIT licence (jobdork/data/humanizer/LICENSE).