JobFiller

Fill job applications from your own CV, without letting an AI make things up.

JobFiller is a self-hosted Firefox extension and a small Django API. Open a job posting, click Scan, check what it plans to write in each field, then click Fill. It handles the fiddly parts: custom dropdowns, forms inside iframes, React-controlled inputs, résumé uploads, cover letters, and multi-page LinkedIn Easy Apply.

Why use it

• Built on your own data. Every answer comes from your CV, the job posting, or a Settings answer you typed. If it can't back up an answer, it skips the field and tells you why.
• It won't click some things for you. It never auto-fills legal attestations, EEO/demographic questions it has no typed answer for, or logistics a CV can't answer (salary, visa, notice period, relocation).
• You check it before sending. Scan and Fill are two separate clicks, and the manual flow never submits the form.
• Works on real ATSs. Tested by hand on Greenhouse, Ashby and LinkedIn Easy Apply.
• Self-hosted. Your CVs stay in a local SQLite database. The extension talks only to your own core instance, never to an AI model directly.
• Extras: tailored CV generation (rendered to PDF with LaTeX), per-scan cover letters, a job fit analysis backed by web search with sources, and an optional Telegram job-channel agent.

How it's built

• core/: Django 5 + DRF API. CV storage and parsing, the job-posting scan, and every AI call. AI calls go through the OpenAI SDK to Xiaomi MiMo by default. You can set MIMO_BASE_URL to another OpenAI-compatible endpoint, but only MiMo has been tested. Without a key, it runs on heuristics alone.
• extension/: Firefox (Zen) WebExtension, Manifest V3. Reads the form on the current page, sends it to core, and applies the plan that comes back.

CLAUDE.md covers the architecture and the traps. tasks/PROGRESS.md lists what's built and why, and tasks/BACKLOG.md lists what's next.

Quick start

Backend, straight from the host:

cd core
cp .env.example .env   # edit SECRET_KEY before any real use
uv sync
cd src && uv run python manage.py migrate
uv run python manage.py runserver 0.0.0.0:8000

Or via Docker (still needs core/.env — same cp step, from core/):

docker compose up --build

core/.env is the only place secrets live. MIMO_API_KEY empty is a supported state: the LLM passes no-op and you get the heuristic plan. TAVILY_API_KEY empty disables Analyze.

Extension, with core on localhost:8000:

1. about:debugging#/runtime/this-firefox → Load Temporary Add-on… → extension/manifest.json.
2. A small tab appears on the right edge of every page — that's the panel. Open it and click Manage CVs to upload a CV, fill in the EEO answers you want used, and generate tailored CVs.
3. On a job posting (a job-boards.greenhouse.io listing is a good first test), open the panel, pick a CV and click Scan & Fill — the same button reads Fill application once the scan comes back, so filling is a second, deliberate click.
4. Read the Logs tab for what was skipped and why, and check the form yourself before submitting. The manual Scan & Fill flow does not submit the form.

There's no toolbar icon and no per-site permission prompt: the panel is a content script and <all_urls> is a required host permission, so installing asks for "Access your data for all websites" once.

Greenhouse, Ashby and LinkedIn Easy Apply have all been driven by hand; tasks/PROGRESS.md marks what's confirmed live versus only reviewed.

On LinkedIn, select a CV and click Fill Easy Apply steps, or use the regular Scan button followed by Fill & continue Easy Apply. The extension opens Easy Apply if needed, fills each page, and presses Continue/Review until the final review page. It leaves Submit application for you to review and click. If the CV cannot support a required answer, it fills the other fields and pauses with the dialog open. Enter your answer and click Continue Easy Apply to resume. This flow has been checked against a multi-page Firefox fixture; it still needs a live LinkedIn run after reloading the extension. If your CV omits a required contact number, enter the phone number and country code in LinkedIn before continuing. LinkedIn has separate country-code and mobile-number fields; if you store a full +7… answer in Settings, JobFiller removes the selected +7 prefix before filling the mobile-number box.

Generating a tailored CV

Manage CVs → Generate a tailored CV rewrites a stored CV for a position you paste in and saves the result as a new PDF you can pick when filling. It typesets with pdflatex, so that one feature needs TeX on whatever runs core. The Docker image installs it. From the host, on Fedora:

sudo dnf install texlive-scheme-medium texlive-fontawesome5 texlive-charter texlive-paracol

Without TeX that endpoint returns a 503 naming what to install; nothing else is affected.

Zen/Firefox installed via Flatpak (common on Linux)

The add-on will load with no error and then do nothing — no corner tab, and a blank Manage CVs tab. The Flatpak sandbox only grants access to the one file the "Load Temporary Add-on…" picker pointed at (manifest.json); every sibling file resolves empty. Fix:

flatpak override --user --filesystem=/absolute/path/to/job-filler app.zen_browser.zen

Then fully quit and relaunch Zen and load the add-on again — Reload alone won't pick up the grant.

What it fills, and what it refuses to

Field mapping runs a heuristic pass first: contact fields (name, email, phone, LinkedIn, GitHub) and the résumé upload come free from regex against your CV. Whatever it skips goes to MiMo in one batched call, grounded in your CV text and the scanned posting, including custom JS comboboxes (Greenhouse/Ashby-style pickers that aren't real <select>s — the extension opens the widget at fill time and picks from the real options, with a second round trip to core when core's blind guess matches nothing).

Never answered, regardless of confidence:

• Legal attestations ("I agree…", privacy and terms consent). Your click to make, not ours.
• Logistics a CV can't answer — travel, relocation, salary, notice period, visa, start date. These are filtered out of the model call entirely rather than left to the prompt.
• EEO/demographic questions, which are never guessed from a CV or posting. Type your own answer once in Manage CVs → Settings and a dedicated pass picks or words the value per field, grounded strictly in what you typed — no CV text, no posting text in that prompt. A blank row stays skipped.

Also there: a cover letter generated per scan from your CV and the posting (text for a paste field, a .docx for an upload field), a per-field Generate with AI button on open-ended textareas, and Analyze Application — a fit score plus company and job-market notes, each grounded in live Tavily search results and carrying its source URL, because a model with no internet would simply invent them.

Generated CVs list back every technology that wasn't already in your source CV, so you can confirm each one is true before sending it anywhere.

Telegram opportunity agent

The proactive service reads a Telegram job channel (set TELEGRAM_CHANNEL_TITLE; the default is Digital nomads. Work from anywhere) from a Telegram account that has already joined it. Get an API ID and hash at my.telegram.org, set TELEGRAM_API_ID and TELEGRAM_API_HASH in core/.env, and log in locally (enter the phone, code, and any 2FA password in the terminal):

cd core/src
uv run python manage.py migrate
uv run python manage.py telegram_login
uv run python manage.py sync_telegram_jobs

Do not paste Telegram secrets or login codes into chat. The account session is stored in core/src/data/telegram.session; keep this file private and back it up with the data directory. The running installation uses Docker: core serves the API on localhost, and telegram-sync checks for new posts every six hours. Both share the core_data volume. For a fresh Docker setup, run docker compose run --rm --no-deps telegram-sync uv run --no-sync python manage.py telegram_login once, then docker compose up -d --build. The Docker volume's CV database and Telegram session are separate from the host core/src/data; upload CVs to the running Docker API or migrate them first.

The service follows each post's Telegraph category links to individual job descriptions and final application links. MiMo first selects promising job titles, then compares their descriptions with uploaded CVs in one batch. Set MIMO_API_KEY to enable this matching. The default MIMO_MODEL and MIMO_VISION_MODEL are mimo-v2.6-flash. The extension sends a screenshot with scanned form fields so the vision model can read visual labels; unsupported controls still need DOM support to be filled. Matches above OPPORTUNITY_MIN_SCORE (default 75) enter a queue. While Firefox/Zen and the extension are running, it polls this queue every 30 minutes, opens one job at a time, scans and fills its form with the selected CV, and submits only when required fields are answered, filling succeeds, and a single final submit button is found. It marks the job applied only after detecting a confirmation; otherwise it leaves the tab open and marks it needs review. For LinkedIn Easy Apply links, the queue worker uses the same multi-page flow and submits after the final review page only if every step succeeds. Reload the Firefox/Zen extension after updating it. See Manage CVs → Telegram opportunities, GET /api/v1/opportunities/, or Django admin for the queue and reasons. Posts without an individual application link stay available for review.