Skip to content

About

One-click local app that runs a team of Claude Code agents in your own Chrome to apply for jobs honestly. Your data stays on your machine. Practice mode, live logs, no fabricating.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Jobs Express

Set it up once. Press Start, or let it run itself every morning. Parallel Claude agents fill out job applications in your own Chrome, and only ever tell the truth.

A free, local, one-click web app that puts Claude Code to work on your job search. Your data stays on your computer. Anything an agent cannot answer honestly is held for you instead of guessed.

Want it on your computer? Install in three steps. Nervous about it? Read Responsible use and limits first. It is short, and it matters.

Tests License: MIT Node 18+ Platform: Windows, macOS, Linux No dependencies


What it does

Jobs Express in action

The recording above is the app's demo mode, so every name, company, and number in it is made up. It is rendered with node tools/record_demo.mjs, and you can open the same sample data yourself with node server.mjs --demo.

You answer a handful of questions once: who you are, what work you want, and the plain facts about your experience. Then you press Start applying. Up to three Claude Code agents open tabs in the Chrome you already use, find roles that fit, fill in the application forms from your answers, and report back in a live log and a table.

The part that makes it different is what the agents will not do:

  • They only claim facts you gave them. No invented numbers, degrees, certifications, or stories.
  • When a form asks something your profile does not cover (a driver's license, a drug-test consent, a criminal-history question), they stop and ask you. Your answer is saved, so you are asked once.
  • In Practice mode they fill every form and stop before the final Submit, so you can look first.
  • They never create accounts, type passwords, click CAPTCHAs, read your email, or message anyone for you.

Once you trust what Practice mode does, you can set a daily schedule: pick a time, and every day the app opens Chrome if needed, looks only at roles posted in the last day or so, applies to every honest fit, and leaves a short digest of what it did. See Run it every day.

Why not just use an auto-apply bot?

Plenty of paid services will blast out hundreds of applications a week. Some are fine. Many answer questions by guessing, which means they lie on your behalf, and the lie is attached to your name. This project makes a different trade: slower, smaller batches, and nothing submitted that you could not stand behind.

Jobs Express Applying by hand Typical cloud auto-apply bot
Cost Free. Uses your own Claude plan Free, but it costs hours Often $30 to $100 a month
Where it runs Your computer, your Chrome You Their servers
Your data Stays in one folder on your machine Wherever you type it Uploaded to a vendor
Honesty Refuses to fabricate. Unknown answers are held for you As honest as you are Frequently fills gaps with guesses
Volume Small batches on purpose (1 to 3 agents) A few a day Hundreds a week
You can review first Yes, Practice mode Yes Rarely
Speed Minutes per application 15 to 30 minutes each Fast, which is part of the risk
Account risk Real. See limits None Real, and out of your hands
Setup About 10 minutes, once None Sign-up and a card

Highlights

  • One click to start. Double-click Start.bat (Windows), start.command (macOS) or run ./start.sh (Linux). It checks Node, starts a small local server, and opens the app in your browser.
  • A daily schedule. Set a time once. Windows Task Scheduler, macOS launchd or a Linux systemd timer starts the run for you, even if this app is closed. A digest lands in the app after every run.
  • Claude Code can be installed and signed in from inside the app. Two buttons in the requirements step run Anthropic's official installer and claude auth login, after showing you the exact command. No API keys: you sign in with your normal Claude subscription.
  • A setup wizard you only do once. Contact details, resume, the roles and pay you want, and an honest list of the facts agents may claim. Everything is editable later in Settings.
  • Requirements checks that explain themselves. Is Claude Code installed? Signed in with the right kind of login? Is Chrome connected? Each row says what is wrong and exactly how to fix it.
  • Practice mode and Live mode. Practice fills forms and stops before Submit. Live submits.
  • One to three agents. Start with one. More agents finish sooner but share a single Chrome and make LinkedIn more likely to notice.
  • A live log you can read. Plain sentences by default, with a raw view when you want to debug.
  • A "Needs your input" inbox. Questions the agents could not answer honestly, with a box to answer once, and a one-click re-run for the postings that were waiting.
  • An applications table. Company, role, pay, source, status, link, and time. Filter it, search it, export it to CSV.
  • No dependencies, no build step. Plain Node and plain browser JavaScript. Read the whole thing in an afternoon.
  • Light and dark themes, keyboard friendly, and a proper phone layout (bottom tab bar, tables become cards).

A closer look

Setup and requirements Your honest profile
Setup: the warning and requirement checks Your honest profile
A run in progress Applications and their status
The dashboard during a practice run The applications table
The daily schedule A morning digest
Daily schedule settings A daily digest, opened
Install Claude Code from inside the app On a phone
The install confirmation shows the exact official command The schedule on a phone
Dark theme
Dark theme

All sample data, all made up.


Install in three steps

  1. Download. Click the green Code button, then Download ZIP (or grab a ZIP from Releases) and unzip it anywhere. Or git clone https://github.com/DatafyingTech/Jobs-Express.git.
  2. Double-click the launcher for your computer.
    • Windows: Start.bat. If Node.js is missing it offers to install it with winget.
    • macOS: start.command. If macOS refuses, right-click it, choose Open, then confirm. On macOS 15 (Sequoia) and newer that shortcut no longer works: try to open it once, then go to System Settings, Privacy & Security and choose Open Anyway. If nothing happens, run chmod +x start.command once in Terminal. (Not tested on a real Mac by the author.)
    • Linux: ./start.sh.
    • Anywhere: npm start or node server.mjs.
    • Windows may show a blue "Windows protected your PC" (SmartScreen) page for a downloaded ZIP. Choose More info, Run anyway if you downloaded it from this repository.
    • Extract the whole ZIP first (right-click, Extract All). Running Start.bat from inside the ZIP window does not work, and the launcher says so.
  3. Follow the wizard. Your browser opens to the app. The first screen checks your computer and walks you through anything missing. Then answer the questions, run a Practice batch, and look at what it did before you ever switch to Live.

The app lives at http://127.0.0.1:4317 (or the next free port) and is only reachable from your own computer.

Want to look around first? Run node server.mjs --demo (or add ?demo=1 to the address). It loads a fictional person, fourteen sample applications, and a scripted fake run. Nothing real is touched and no browser or Claude call is made.


Requirements

You need
Node.js Version 18 or newer (nodejs.org, the LTS release is fine). The tests run on Node 18, 20 and 22 in CI
Claude Code Installed and working. You can install it from inside the app: the requirements step has an Install Claude Code button that runs Anthropic's official installer (it shows you the exact command and its source, the quickstart, and runs only when you press the button)
A claude.ai plan Pro, Max, Team or Enterprise. No API keys needed: you sign in with your normal Claude subscription, from inside the app with the Sign in to Claude button (it runs claude auth login --claudeai and opens your browser). API-key logins do not enable the Chrome integration
A browser Google Chrome or Microsoft Edge
The extension Claude in Chrome, version 1.0.36 or newer, installed and signed in to the same claude.ai account
LinkedIn If you use LinkedIn Easy Apply: signed in, by you, in that same Chrome profile
OS Windows 10/11, macOS, or desktop Linux. WSL is not supported by the Chrome integration

The app checks all of this on the first screen and tells you what to fix. You do not need to memorize it. The Chrome extension step stays manual on purpose: install it from the Chrome Web Store (the app has a direct link), sign in, and press Test connection.


Run it every day

Settings, Daily schedule. Pick a time, whether it is every day or weekdays, Practice or Live, how many agents (1 to 3), the most applications per run (default 10, hard maximum 25), how fresh a posting has to be (default 30 hours), and a time cap (default 90 minutes). Press Save schedule. The app registers a job with your operating system that runs node server.mjs --scheduled-run at that time:

System What is created If the computer was off or asleep
Windows A Task Scheduler task named JobsExpressDaily, run only while you are signed in Starts when the PC wakes (StartWhenAvailable, and it may wake the PC)
macOS A launchd agent in ~/Library/LaunchAgents launchd runs it when the Mac wakes. A Mac that was shut down does not catch up
Linux A systemd user timer with Persistent=true, or a crontab line if systemd user timers are not available systemd catches up at next start. Cron does not

What has to be true for it to run, honestly:

  • On Windows, Claude Code also wants Git for Windows (Git Bash). The requirements step warns if it is missing.
  • The computer is on and you are signed in to it (a locked screen is fine). It drives your own desktop Chrome, so a signed-out or shut-down machine cannot run it. On Windows a console window shows while it runs. Leave it alone.
  • Chrome has the Claude extension installed and signed in, and LinkedIn is signed in in that profile. The run starts Chrome if it is closed, but it never signs in for you. If a prerequisite is missing it writes a "did not run" digest that says what to fix, and exits.
  • Only one run at a time: a scheduled run skips itself while you are running something by hand, and the other way round.
  • Practice is the default. Live needs a tick in a box that says "I have reviewed Practice results and accept that applications will be submitted automatically". Start with Practice, read a few digests, then decide.
  • The freshness window is not a quota. Only roles posted in the last N hours are considered, and the agents apply to every honest fit and never pad the run with weaker matches. Zero applications is a normal result.
  • If you move this folder or update Node.js, press Save schedule again. The Status box warns you when the job is out of date.

Run a test now starts a small Practice run (up to 3 applications, 30 minutes) through exactly the same steps, so you find problems at 2 p.m. instead of 4 a.m. From a terminal, node server.mjs --scheduled-run --dry-run prints what a run would do and what would stop it, without launching anything (add --json for machine-readable output).

Every run leaves digests/YYYY-MM-DD.md and .json in your data folder, and the same list appears in the app under Daily digests. Anything an agent could not answer, including a site that emailed a verification code, appears under Needs your input with the company, the role and a link, so you can finish it yourself. The agents never read your email.


How it works

flowchart LR
    You([You]) -->|answer once| Wizard[Setup wizard]
    Wizard -->|saves| Profile[(profile.json<br/>in your data folder)]
    You -->|Start applying| Server[Local server<br/>127.0.0.1 only]
    Sched[Task Scheduler, launchd<br/>or systemd timer] -->|once a day| Server
    Server -->|builds prompts from| Profile
    Server -->|spawns 1 to 3| Agents[claude -p --chrome<br/>one process per agent]
    Agents -->|drive| Chrome[Your Chrome<br/>with Claude in Chrome]
    Chrome --> Sites[LinkedIn, company job boards]
    Agents -->|node tools/log.mjs| Log[(applications.jsonl<br/>held.jsonl)]
    Log --> Dash[Dashboard:<br/>table and inbox]
    Agents -.->|live stream| Dash
    Dash -->|your answers| Profile
Loading
  1. Your answers are saved to profile.json in your data folder. That file is the only source of facts the agents may claim.
  2. When you press Start, the app writes one prompt per agent: shared honesty rules, a guide for each source, your profile, and a mode block (Practice stops before Submit). The prompt is sent to Claude Code on standard input, so it never appears on a command line.
  3. Each agent is claude -p --chrome with a deliberately small toolbox: a named list of Chrome tools (no page scripts, no screenshot uploads), WebFetch for the public job feeds only, WebSearch, and a single allowed shell command, node tools/log.mjs. There is no file-reading or file-writing tool. It never runs with --dangerously-skip-permissions, it never prompts (anything not on the list is denied), your own Claude settings and hooks are ignored for these runs, your other connectors (mail, drive, chat) are switched off, and the agent only inherits an allow-list of environment variables, never tokens or keys. Each agent also runs under a small guard that stops it if the app is closed or killed.
  4. Agents record every outcome with tools/log.mjs. The dashboard reads those files and streams the agents' progress live.
  5. Questions an agent cannot answer land in Needs your input. Your answer is saved into the profile's memory and used word for word from then on.

The prompts are plain Markdown in agents/. Read them. They are the heart of the project, and you are welcome to improve them.


Responsible use and limits

Please read this part. It is why the app puts a warning in front of you before your first run.

  • LinkedIn's terms restrict automation. Using this on LinkedIn can get your account limited or restricted. That risk is real and it is yours. Start with small batches, use Practice mode first, keep to one agent until you are comfortable, and stop immediately if LinkedIn shows warnings.
  • You are responsible for what is submitted in your name. The agents are built to refuse to fabricate, but you are the one who wrote the honest profile, and you are the one who should read the Practice results. If your profile overstates something, the agents will repeat it.
  • Be signed in yourself. You sign in to LinkedIn, in the same Chrome profile that has the Claude extension. The extension has to be installed and signed in too. The app never asks for, sees, or stores your passwords.
  • Unattended runs are still your responsibility. A scheduled Live run submits applications while you are away. Read the digests, and switch back to Practice the moment anything looks off.
  • Do not use that Chrome profile during a run. The agents share it, and your clicks can collide with theirs. Keep the Chrome window open until the run finishes.
  • What the agents will never do: create accounts, set or enter passwords, click CAPTCHAs, enter emailed verification codes, invent references, accept binding arbitration or NDA boxes, read your email, or message or post for you. When they hit one of these they hold the application and tell you.
  • Respect the sites. Agents back off on rate limits. Do not edit the prompts to hammer a site.
  • Some employers say "no AI-written applications". The agents are told to skip those. Check yourself that you are comfortable with how an application was prepared.

This is a personal tool, not a service, and it comes with no warranty (see LICENSE).


Privacy

  • Your data is a folder, outside the app folder. On Windows it is %LOCALAPPDATA%\JobsExpress, on macOS and Linux ~/.jobs-express (or wherever CJA_DATA_DIR points). It holds your profile, a copy of your resume, the applications table, the held questions, digests and run logs. Keeping it out of the app folder keeps the agents, which work inside the app folder, away from it. If you used 1.0, its data/ folder is moved there automatically the first time you start 1.1. To delete everything, close the app and delete that folder.
  • This app has no servers, no account, no telemetry, and no analytics. The local server only listens on 127.0.0.1. It answers only requests for its own address and port (so a web page cannot reach it through DNS rebinding), only accepts changes that come from its own page, and gives a session only to the window the launcher opens: the launcher opens a one-time link that sets a private cookie, and the page itself contains no secret. A website you visit, or another program or user on this computer, therefore cannot start agents or read your data by just asking the server for its page. A folder the app creates is private to your user account (mode 0700 on macOS and Linux, a restricted ACL on Windows); an existing folder you point it at is left alone and the app tells you if other accounts can read it.
  • What leaves your computer, and where it goes:
    • Claude Code sends your prompt (which includes your profile) and what the agent reads in the browser to Anthropic, as it does for any Claude Code session. See Anthropic's data usage page. Test connection, and the connection check before each scheduled run, also make a tiny real Claude request.
    • The agents can use WebFetch (only the public job-board feeds and Hacker News) and WebSearch, so search terms such as job titles, and the addresses of those feeds, go out too. They are told never to put your personal details in a search or a URL.
    • The Chrome extension acts on the sites you ask it to, using your signed-in sessions.
    • The job sites you apply to receive the application, including your resume.
  • Nothing else. The app does not upload your profile or resume anywhere itself.
  • A note on screen recordings. The agents' log shows what they typed into forms. If you record or screenshot it, check it first.

Troubleshooting

What you see What to do
"Claude Code was not found" Press Install Claude Code in the requirements step (or install it from the quickstart, open a new terminal, run claude --version), then press Re-check
The daily run did nothing Open Settings, Daily schedule: the newest digest says what stopped it (Chrome's extension signed out, LinkedIn not confirmed, Claude signed out, the computer was off). scheduled-run.log in your data folder has the full text
The schedule says it is out of date You moved the folder or updated Node.js. Press Save schedule again
"Claude Code is using an API key" The Chrome integration needs a claude.ai login. Unset ANTHROPIC_API_KEY, run claude auth login, choose the claude.ai option. The app also strips API-key variables from the agents' environment
Test connection fails Install the Claude extension (1.0.36+), open it and sign in with the same account, close every Chrome window and open Chrome once more (it reads its connection settings at startup), then test again
"Receiving end does not exist" or it works then stops The extension's background worker went idle. In Claude Code run /chrome and choose Reconnect extension, or toggle the extension off and on in chrome://extensions
Agents stall or nothing happens A pop-up dialog in Chrome may be blocking it. Dismiss it and press Stop, then start again
LinkedIn shows a security check Handle it yourself in Chrome, wait a while, and use fewer agents. Never let anything automate a CAPTCHA
Windows says "Node is not recognized" Install Node.js LTS (the launcher offers winget), then close the window and open the launcher again
macOS will not open start.command Right-click, Open. Or in Terminal: chmod +x start.command
Port in use The app picks the next free port automatically and prints the address
A run stopped by itself Runs have a time cap (45 minutes) and an action cap per agent (200). Change them in Settings, Run limits
Where are the logs? runs/<run id>/ in your data folder has a readable run.log and each agent's raw output. The newest 20 runs are kept
"This window was not opened by Jobs Express" The page needs the one-time link the launcher opens. Close the tab and start the app again. If it is already running, starting it again opens the right window

FAQ

Does it work without LinkedIn? Yes. Turn off LinkedIn Easy Apply in the wizard. The agents then work from company job boards (Greenhouse, Ashby, Lever and similar) and aggregators.

Will it apply to everything? No. It skips roles under your pay floor, anything matching your exclusions, roles that need what your profile says you do not have, and companies you have already applied to.

What happens when it hits a question it cannot answer? It holds the application, writes the question to Needs your input, and moves on. Answer once, tick "remember", and it will not ask again.

Can I see exactly what the agents are told? Yes. Read agents/rules.md and the other files in agents/. lib/prompts.mjs assembles them.

Does it cost anything? The app is free. Agents use your Claude plan's usage, and a run of several applications is a real amount of it. Start with a small batch.

Why only three agents? They share one Chrome window and one LinkedIn account. More would mostly mean more collisions and a higher chance of being rate limited.

Can I use another browser? Chrome and Edge are the supported ones. Claude Code can also work with some other Chromium browsers, but this app only checks for Chrome and Edge.

Is it safe to run --dangerously-skip-permissions? The app never does, and neither should you here. The agents get a short allow-list of tools instead.

Does it send emails or messages? Never. It does not read your email either. If a site wants an emailed code, the application is held for you: it shows up in Needs your input with the company, the role and a link, and you enter the code yourself.

Do I need an API key? No. You sign in with your Claude subscription (Pro, Max, Team or Enterprise). API keys do not work with the Chrome integration, and the app removes them from the agents' environment.

Can it run while I am away from the computer? Yes, if the computer is on and signed in. See Run it every day for the honest list of conditions.

How do I update it? Download the new ZIP (or git pull) and start it. Your data lives outside the app folder, so there is nothing to copy. If you ran 1.0, its data/ folder is moved out of the app folder for you the first time.


Development

node --test            # run the tests (no install step)
node server.mjs --demo # the UI with sample data
node server.mjs --scheduled-run --dry-run   # what a daily run would do, launching nothing
node tools/record_demo.mjs   # re-render docs/images (Playwright or headless Chrome/Edge, plus ffmpeg for the GIF)
node tools/check_layout.mjs  # needs Playwright: fails if any page scrolls sideways at 360, 390, 768 or 1440 px

See CONTRIBUTING.md. Security issues: SECURITY.md. Changes: CHANGELOG.md. A longer walk-through is in docs/guide.md.

Credits

Built by Datafying Tech on top of Claude Code and the Claude in Chrome extension. The honesty rules came out of a lot of real job applications and a lot of near misses. MIT licensed.

About

One-click local app that runs a team of Claude Code agents in your own Chrome to apply for jobs honestly. Your data stays on your machine. Practice mode, live logs, no fabricating.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages