Roleflow

An open-source AI agent that finds real jobs, tailors your resume to each one, and fills in the actual application in your browser. Nothing is submitted until you approve.

Demo · Features · Quick start · Architecture · Contributing

Demo

▶ Watch the full demo (75 s, MP4) · made with Remotion

_{Real UI and a real agent run on a live Greenhouse posting with a fictional candidate. The run was stopped at the approval step, so nothing was submitted.}

Why

Applying for jobs is the same loop over and over: find a role, tweak your resume, retype the same answers into another form, upload, repeat. Roleflow automates the loop but keeps you in charge:

• It works on live listings and the employers' real application forms, not a sandbox.
• It never invents experience, dates, salary or work authorization. When it doesn't know something, it asks you.
• It stops before the final Submit and shows you exactly what it filled.

Features

🔎 Live job search	LinkedIn's public listings plus company career sites on Greenhouse, Lever and Ashby (found with OpenAI web search and verified through each ATS's public API). Results are ranked against your resume. You can also paste any job link.
📄 Resume understanding	Upload a PDF or DOCX. The text and hidden hyperlinks are parsed into a structured profile.
🎯 Match analysis	A score with concrete evidence from your resume, the gaps, and the things worth confirming before you apply.
✍️ Per-job tailoring	A rewritten resume PDF and cover letter for each job. Every edit is listed with a reason. Employers, titles, dates and education are restored in code, and skills you never listed are dropped.
🤖 Browser agent	Opens the real application in Chrome and fills it in: text fields, dropdowns, type-aheads, checkboxes and resume uploads. Multi-step and LinkedIn Easy Apply flows are supported. A live view and timeline show every action.
🙋 Human in the loop	Asks for missing facts (and remembers your answers), pauses for sign-ins and CAPTCHAs, and waits for Approve & submit.
🚀 Autopilot	Give it a role, a location and a number of applications. It searches, picks the best matches, tailors your resume for each and applies one by one, still stopping for your approval.
🛡️ Guardrails in code	Submit gate, browser-validation check, demographic "decline" enforcement, URL allow-list, retry and token caps.

Screenshots

Live search + match analysis. Real listings ranked by fit, with evidence and gaps.	Browser agent. It filled a real Ashby form, uploaded the tailored resume and is waiting for approval.
Tailored resume. Each change explained; anything not on your resume is listed as not claimed.	Autopilot. Runs from search to submit for several jobs, with a readiness checklist.
Workspace. Natural-language search, sources, and the parsed resume profile.

_{Screenshots use a fictional candidate.}

How it works

flowchart LR
  R["Resume PDF/DOCX"] --> P["Parse<br/>gpt-5.5 structured output"]
  Q["Role + location"] --> S["Search<br/>LinkedIn · Greenhouse · Lever · Ashby"]
  P --> M["Match + tailor<br/>resume PDF + cover letter"]
  S --> M
  M --> A["Browser agent<br/>observe → decide → act"]
  A --> H{"Needs you?"}
  H -- "question / sign-in" --> U["You answer or act"] --> A
  H -- "form complete" --> V["Approve & submit"] --> D["Application submitted"]

Loading

Each agent step sends the model a numbered map of the page's interactive elements (across iframes, shadow DOM and modals), a low-detail screenshot, your facts and a compact action history. The model replies with tool calls (fill_fields, upload_file, click, ask_user, ready_to_submit…), which a Playwright executor carries out and verifies.

The full design, with data flows, state machine, guardrails, data model and API, is in docs/ARCHITECTURE.md.

Quick start

Requirements: Node.js 20.9+, Google Chrome (or Edge, via BROWSER_CHANNEL=msedge), and an OpenAI API key.

git clone https://github.com/Deveshu04/roleflow.git
cd roleflow
npm install
cp .env.example .env.local      # then set OPENAI_API_KEY
npm run dev

Open http://127.0.0.1:3000. The landing page is at / and the workspace at /app.

1. Upload your resume, then open Application details and fill in your location, work authorization, notice period and salary expectations, so the agent rarely has to stop and ask.
2. (Optional) Click Connect LinkedIn and sign in yourself in the Chrome window that opens. The session stays in the agent's own browser profile, and the app never sees your password.
3. Search for a role or paste a job link, then Analyze, Tailor for this job and Run application agent. Or switch to Autopilot to do several jobs in a row.
4. Answer any question in the agent panel, check the filled form in Chrome, and click Approve & submit. Don't submit stops the run and leaves the tab open.

Configuration

All settings are environment variables in .env.local:

Variable	Default	Description
OPENAI_API_KEY	—	Required. Used for parsing, matching, tailoring, web search and the agent.
OPENAI_MODEL	gpt-5.5	Model for parsing, matching, tailoring and search.
OPENAI_AGENT_MODEL	OPENAI_MODEL	Model that drives the browser agent.
BROWSER_CHANNEL	chrome	chrome, msedge or chromium.
HEADLESS	false	Hide the agent's browser. You then can't sign in or solve CAPTCHAs.
AGENT_MAX_STEPS	40	Model steps per application.
AGENT_MAX_RUN_TOKENS	150000	Input-token budget per application; the run stops and leaves the tab open.
ROLEFLOW_DATA_DIR	./data	Where resumes, runs and the agent's browser profile are stored.

Cost: each agent step is about 3–6k input tokens. On live Greenhouse, Lever and Ashby forms a full application took 8–13k input tokens. Every run shows its token usage.

Privacy and responsible use

• Local-first. Everything stays in data/ on your machine, which is git-ignored. The server binds to 127.0.0.1 only, and the API rejects cross-site and non-localhost requests. Resume text and job pages are sent to OpenAI for processing.
• Your credentials stay yours. Roleflow never asks for passwords, never creates accounts and never bypasses CAPTCHAs or bot checks.
• You are the applicant. Review every form before approving. Auto-submit is off by default.
• LinkedIn: automating actions on your LinkedIn account may break LinkedIn's User Agreement and could lead to restrictions. Keep the volume low and apply only to roles you'd apply to by hand.

Development

npm run dev          # http://127.0.0.1:3000
npm run check        # lint + typecheck + tests
npm run build        # production build
npm test             # agent form-engine tests (no API key needed)

The tests run the agent's page reader and form filler against fixture forms modelled on Greenhouse, Lever and Ashby, including Lever's location type-ahead, hidden checkboxes, comboboxes, validation and the demographic guard.

CI runs on every push and pull request:

• lint, typecheck and tests on Node 22 and 24
• a production build
• gitleaks secret scanning and npm audit
• dependency review
• CodeQL

app/            workspace UI and API route handlers
lib/jobs        job sources, import, ranking, match analysis
lib/resume      parsing, tailoring, PDF rendering
lib/agent       browser agent: page snapshot, actions, runner, autopilot
public/landing  landing page
tests/          form-engine tests
video/          Remotion source of the demo video
docs/           architecture, screenshots, media

Contributing

Contributions are welcome. Please read CONTRIBUTING.md and the Code of Conduct, and report security issues privately as described in SECURITY.md.

License

MIT © 2026 Deveshu Pathak

_{Built with Next.js, the OpenAI API, Playwright and Remotion. Roleflow is not affiliated with LinkedIn, Greenhouse, Lever or Ashby.}