Apply Jobs Automatically — AI Agent

Applying for jobs is a numbers game that nobody has the stamina to play properly. Every serious application means reading the description, deciding whether it is worth it, re-emphasising your CV to match, and then twenty minutes of retyping the same phone number into a form that already parsed your resume and got half of it wrong.

This is five AI agents that do that loop for you, inside Claude Code, in your own browser, with your own logins.

What this actually does

You open a dashboard on your own machine, pick a role and a platform, and click Start Applying. From there:

• job-scout searches your chosen platforms and returns each posting as compact structured data
• gatekeeper runs four gates over every single one and skips most of them, in writing, with a reason
• cv-tailor re-emphasises your CV for the jobs that survive, builds a PDF, and visually inspects it before letting it through
• form-filler fills the application form, uploads the CV, and submits
• digest-writer writes the daily summary and updates your tracker

Everything is a file on your computer. There is no service to sign up for, no data leaving your machine, and no company holding your CV. The dashboard is a small Node server on localhost:4747 with zero dependencies.

The honest numbers

A good day is 10 to 15 real submissions. Not 100.

Every tool in this category implies the opposite, so here is where the other 85 go. In one real run, roughly 38,000 harvested postings produced about 50 that were worth applying to. The losses are not the software failing — they are the software refusing:

• Postings that require work authorisation you do not hold
• Recruiter listings with no named end client, which are usually fishing for a CV database
• Roles whose core stack is something you would have to lie about knowing
• "Remote" roles that mean remote-within-one-country
• Duplicates of something you already applied to three weeks ago

An agent that submits 100 applications a day is submitting to all of those. That is not more productivity, it is more noise with your name on it.

Why most job-application bots fail

They get blocked, and they get blocked at the last step — after filling the entire form.

Modern applicant tracking systems run real bot detection. A headless browser has a fresh, obviously-synthetic fingerprint, no session history, and no logged-in state. It sails through the form and then hits a challenge at submit, which is the most expensive possible place to fail.

The workaround that actually works is not a better headless browser. It is not using one for the submit. This pipeline drives your real Chrome, with your real profile and your real logins, through the Claude in Chrome extension — and falls back to fast headless automation only for reading public pages, where nothing is defended.

Knowing which platform needs which tool is the difference between a run that submits and a run that wastes an hour. That map took real applications to build, and it ships with the project.

Which platforms fight automation — and how it gets through

Two details that each save an entire application:

Upload the resume first, then fill the fields. On both Ashby and Greenhouse the upload re-renders the form and silently wipes everything already typed. Do it in the other order and you will submit a form with three empty required fields.

Never follow to.indeed.com redirects. Find the company's own application URL instead.

This map lives in the project as a file the agent reads before touching a browser, and it updates as it learns. If you see it switch tools mid-run, that is the design working.

The four gates

Every posting passes through four checks before a single character is typed. Each one is terminal — there is no override, no "apply anyway", no confidence threshold.

Gate 3 is the one people underestimate. The failure mode of every AI CV tool is quietly adding a skill because the job asked for it. This one is structurally forbidden from doing that: it may only re-emphasise what is already true in your CV, and the list of things it must never claim is yours to write.

What it will never do

These are not settings. They are rules in the file every agent reads on every run.

• Never completes a CAPTCHA or any bot check. It stops and tells you.
• Never creates an account and never enters a password.
• Never enters bank, card, government-ID or tax details.
• Never fabricates an employer, a date, a degree, a certification or a metric.
• Never submits a job you did not authorise.
• Never runs LinkedIn unattended. Supervised only, and it stops instantly on any challenge.

The point of an agent applying on your behalf is that everything it sends is something you would be comfortable defending in an interview. An agent that embellishes to raise its hit rate is worse than no agent.

What you need

About 20 minutes to set up, most of it installing prerequisites. Then roughly 15 minutes a day.

What's inside the setup guide

What's included

• The complete walkthrough, readable on this page - nothing to download
• Five copy-paste prompts that build and configure the whole system
• All 28 project files in full, each with a copy button
• The Pro / Max model switch, so it fits the Claude plan you already pay for
• The CV prompt - hand it your existing CV, or let it interview you and write one
• Lifetime updates as the pipeline improves