Job Application Agent

An automation tool for the job-application loop: it discovers open roles, tailors a resume to each job description, generates a PDF, and fills out the ATS application form (Greenhouse, Ashby, Lever, and others) — including dropdowns, comboboxes, and open-ended questions.

Demo video: https://www.youtube.com/watch?v=SM7EIgbBiiY

What it does

• Discovers open roles across job boards (discover.py, yc_discover.py)
• Tailors the resume to each job description and generates a fresh PDF (tailor.py, gen_pdf.py, make_cv.py)
• Fills the application in its own browser window, so the operator's main browser stays free
• Survives interruptions: queue processing picks up where it left off
• Follows up: drafts follow-up emails for submitted applications (followup.py)
• Hosted UI (webapp.py): a job-queue page for reviewing discovered postings, plus a live-view auto-apply flow that streams the browser session and lets a reviewer intervene mid-fill

How it works

1. data/profile.yaml holds the candidate's facts and voice notes — the single source of truth for identity, work history, and preferences
2. Playwright opens the job URL in a dedicated Chromium instance with a persistent browser profile, so logins and cookies carry over between runs
3. The form is extracted from the page DOM. Known fields (name, contact info, work authorization, EEO) are filled deterministically from the profile; an LLM drafts answers for open-ended questions
4. The agent fills the form and attempts to submit. The final application state should be reviewed — see "Review before relying on auto-submit" below.

Setup

Local (Python virtual environment)

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
playwright install chromium

Docker

docker compose build
docker compose run --rm agent python3 main.py <job_url>

For the hosted job-queue UI:

docker compose -f docker-compose.dokploy.yml up -d

Candidate data

Create data/profile.yaml with the candidate's details (name, contact info, work history, education, work authorization, EEO answers, and voice notes). This file is gitignored — it is never committed, since it holds personal information. There is no template checked in; see the field references in main.py (apply_deterministic_answers) and tailor.py for the expected schema.

LLM credentials

export LLM_API_KEY=...
# optional:
export LLM_BASE_URL=...
export LLM_MODEL=...

Any OpenAI-compatible endpoint works (OpenAI, Groq, etc.).

Run

python main.py <job_url>
python main.py --queue data/jobs.txt      # one URL per line
./apply.sh                                 # launcher script, same as above
python discover.py --dry-run               # preview matching postings without writing the queue
python webapp.py                           # hosted job-queue + live-view UI on :8765

Review before relying on auto-submit

The agent fills the form and attempts to submit it, including a fallback that bypasses the click handler if the initial submit does not confirm. This is not gated behind a human-review pause. Before pointing this at postings that matter, verify the fields it fills are correct for the candidate in question — in particular:

• Work authorization and EEO fields are filled deterministically from data/profile.yaml. Confirm those values are accurate for the candidate before running against a real posting.
• Open-ended questions are LLM-generated and should be spot-checked.
• Some ATS forms use custom widgets (autocomplete location fields, radio groups for work authorization) that are not always filled reliably — check the run's log output for fields reported as still empty.

Other notes

• Multiple candidates are not yet supported as a first-class feature — the tool currently assumes one operator, one candidate, and one profile. See MULTI_CANDIDATE_PLAN.md for the plan to support a staffing-firm, multi-candidate workflow.
• Do not point this at LinkedIn Easy Apply (Terms of Service) or CAPTCHA-walled application flows.
• Workday: log in once in the agent's browser window; the persistent profile remembers the session on later runs. Full Workday auto-apply is not supported yet — matching postings are routed to a manual-review queue.
• Browser profiles, resumes, and application data (data/, outputs/, .browser-profile/, resumes/) are gitignored and must stay that way — they contain session cookies and personal information.