OrbitApply

AI-powered autonomous job search and application system — runs entirely on your local machine.

OrbitApply uses a coordinated fleet of AI agents to find matching jobs, research companies, tailor your resume and cover letter for each role, and track every application through the full pipeline — from first apply to offer. Everything runs on localhost:3000. No cloud. No subscriptions. No data leaves your machine.

Free. Open source. Built for anyone looking for a job or preparing for an interview.

---

Why I built this

I spent the better part of 20 years running large transformation programs — change management, system overhauls, multi-country rollouts. That was my world for a long time.

Then I decided it was time to transform myself.

That journey wasn't easy. I left a familiar identity and walked into a lot of unknowns. Over the last two years I spent 2,500+ hours learning AI — building production-grade applications, agentic workflows, and automation systems entirely through prompt engineering. No traditional code. Just curiosity, persistence, and a belief that the tools had finally caught up to what was possible.

Along the way, I watched a lot of people going through something I'd never had to think about before — the modern job search. And the more I looked, the clearer it was that the process was broken in ways that weren't the candidate's fault.

Then a friend reached out. Finance and accounting background — sharp, experienced, genuinely good at what they do. But not from tech.

He was spending hours every day just trying to find the right roles. Not applying — just finding. Then more hours rewriting his resume by hand for each one. And somewhere in the middle of all that, losing track of what he'd sent, what needed a follow-up, what had simply gone quiet. He had no system. Just tabs, spreadsheets, and a growing sense that the process was designed to exhaust you.

He tried using ChatGPT and Claude to speed things up. Most people do. But here's the problem with that approach — and the market data makes it clear:

• 75% of resumes are rejected by ATS systems before a human ever sees them, and a general-purpose LLM has no idea how a specific ATS actually parses and scores your file
• Both ChatGPT and Claude default to recognizable house styles — "results-driven," slightly stilted formality — that experienced recruiters notice immediately
• In 2023, using ChatGPT for your resume was a clever hack. In 2025, it's just part of the noise — hiring managers are now attuned to generic AI output and flag it fast
• Using an LLM means starting from scratch every single session. No memory of your profile, no company research, no pipeline tracking, no follow-up reminders. You still have to do all of that manually

100 applications at 45 minutes each is 75 hours — two full unpaid work weeks just filling out forms. An LLM cuts the time per application but doesn't solve the coordination problem. You still have to find the jobs, manage the tracker, prep for each company, and remember who said what when.

I built him a small tool to handle all of that end-to-end. It helped.

So I cleaned it up, made it free, and put it here.

If you're job hunting — this is for you. If you're preparing for an interview — this is for you. If you know someone who needs it — please share it.

No subscription. No account. Runs on your own machine.

---

Screenshots

Dashboard — run the full pipeline with one click

SCOUT Results — ranked job matches with fit scores

Pipeline — track every application through every stage

---

How the workflow works

A single click of Run on the dashboard triggers ORBI, the master orchestrator. ORBI sequences all agents in order, streams live step-by-step progress to the UI, and produces a structured result at the end. Here is exactly what happens, in order.

---

Step 1 — SCOUT: Job Discovery

SCOUT reads your memory/profile.json to understand your target roles, locations, salary range, and skills. It then builds up to 15 targeted search queries across two tiers:

• Primary queries (advanced Tavily depth) — high-signal titles like "Director of AI", "Head of AI", "Agentic AI" scoped to direct ATS platforms: Lever, Greenhouse, Ashby, Workday, SmartRecruiters
• Extended queries (basic Tavily depth) — broader role variants and profile-driven terms

SCOUT runs all queries in parallel, collects raw results, then scores every result locally using a pure JavaScript fit scoring engine — no AI call at this stage, which keeps costs near zero for discovery.

Fit scoring breakdown (0–100):

Factor	Max points
Title match (senior AI signal keywords)	35
Location match (remote preference / city)	25
Salary range overlap	20
Skills keyword match	20

• Results scoring below 40 are dropped silently.
• Results scoring 40–69 are shown in the UI for your reference.
• Results scoring 70+ are marked qualified and passed to the next stage.
• Any company in memory/blacklist.json is filtered out before results reach ORBI.

SCOUT saves today's results to workspace-scout/results-{date}.json.

You can also manually import a job by pasting a URL into the dashboard. SCOUT will fetch the page via Tavily, extract the title and description, score it, and add it to today's results.

---

Step 2 — RECON: Company Intelligence

For each qualified job (up to 8 companies per run), RECON builds a structured intelligence profile using Tavily web searches:

• Employee reviews and culture signals
• Salary benchmarks for the role
• Recent news — funding rounds, layoffs, product launches
• Tech stack from engineering blogs and job descriptions

RECON sends all gathered data to Claude Haiku to synthesise into a structured JSON profile containing: company size, funding stage, tech stack, culture summary, Glassdoor rating, salary benchmark (low/mid/high), red flags, opportunity score (0–100), and risk score (0–100).

Each profile is saved to workspace-recon/{company-slug}-intel.json and passed to TAILOR for document personalisation and to COACH for interview prep.

If RECON hits an API rate limit, it skips remaining companies and continues the run with the intel already gathered.

---

Step 3 — TAILOR: Document Generation

TAILOR processes every job with a fit score of 60 or above. For each job it runs two sequential Claude Sonnet calls:

Resume tailoring:

• Loads your base resume from memory/resume.md
• Loads the job description snippet and RECON intel for the company
• Rewrites bullet points to mirror JD language, injects ATS keywords, and personalises the summary
• Runs an internal ATS simulation — targeting 70+/100
• Returns the tailored resume as Markdown with an ATS score and list of injected keywords

Cover letter writing:

• Writes a 250–350 word cover letter with a clear outcome-first structure
• Tone: professional, direct — no filler phrases like "I'm excited to apply"

Both documents are saved into a named folder:

Apply/applications/{Company Name} — {Job Title}/
├── {Title}_{Company}_Resume_{FirstName}.doc
├── {Title}_{Company}_Coverletter_{FirstName}.doc
└── metadata.json

Before TAILOR runs for any job, GUARDIAN runs a pre-flight check:

• PASS → proceed to tailoring
• BLOCK → company is blacklisted or protected contact — skip silently
• PAUSE → sensitive form fields detected — add to human queue, skip
• HARD_STOP → daily budget or apply cap reached — stop the entire run

---

Step 4 — LEDGER: Pipeline Registration

Every successfully tailored job (ATS score ≥ 70, or the only result) is registered in workspace-ledger/pipeline.json as a new application record with:

• Full metadata: title, company, URL, platform, fit score, ATS score, document paths
• Initial status: applied
• Follow-up reminder: set to 7 days after creation
• Budget cost per application recorded (≈ $0.08 per application)

LEDGER maintains running stats: total applied, response rate, today's count, total budget spent.

---

Step 5 — SUBMIT: Application Submission (optional)

SUBMIT is only active if submit.autoSubmit: true is set in orbitapply.json. It is off by default.

When enabled, SUBMIT uses Playwright to automate form filling on supported ATS platforms:

Platform	Support
Workday	Multi-step application wizard
Greenhouse	Standard form + custom questions
Lever	Application form + cover letter field
iCIMS	Enterprise ATS form fill
Direct career pages	Best-effort Playwright

SUBMIT enforces a minimum 45-second delay (±30% jitter) between submissions to avoid triggering spam detection. It fills standard fields from your profile, pastes the tailored resume and cover letter, takes a screenshot after each form page, and logs the full submission.

GUARDIAN runs again immediately before each submission. If the daily apply limit (15) has been hit, or the daily budget ($5) is exhausted, SUBMIT hard-stops and does not continue.

---

After the run — ongoing pipeline management

Once a run completes, the dashboard shows your full application pipeline board. You move applications through stages manually as real responses come in:

applied → viewed → phone_screen → interview_1 → interview_2 → offer
                                                              → rejected
                                                              → withdrawn

When you update a status to phone_screen, interview_1, or interview_2, LEDGER triggers COACH.

---

COACH: Interview Preparation (triggered automatically on stage change)

COACH activates the moment you mark an application as reaching an interview stage. It loads the RECON intel for that company and generates a full prep pack saved to workspace-coach/{job-id}-prep.md:

1. Company overview — from the RECON intelligence profile
2. Role analysis — what the interviewer is actually testing for
3. 10 behavioural questions with STAR-method answer templates
4. 5–10 technical/domain questions matched to the role
5. Salary negotiation script — anchored 15–20% above the RECON salary benchmark mid-point, with exact counter-offer language
6. 5 strategic questions to ask the interviewer
7. 3 recent news items to reference naturally in conversation

---

Budget and safety controls (GUARDIAN)

GUARDIAN runs as a gating layer before every TAILOR and SUBMIT action. It cannot be bypassed by any agent.

Control	Default	Configurable in
Daily budget cap	$5.00 (hard stop)	orbitapply.json → budget.dailyLimitUSD
Budget alert threshold	$3.00 (warn)	orbitapply.json → budget.alertAtUSD
Max applications per day	15	orbitapply.json → guardian.maxAppliesPerDay
Rate limit between submits	45 seconds	orbitapply.json → guardian.rateLimitMs
Human pause fields	salary, diversity, custom_essay, security_clearance	orbitapply.json → guardian.humanPauseFields

---

Agent summary

Agent	Model	Role
ORBI	Claude Sonnet	Master orchestrator — sequences all agents, streams progress, enforces fit threshold
SCOUT	No AI (pure JS)	Job discovery via Tavily — 15 parallel searches, local fit scoring
RECON	Claude Haiku	Company intelligence — culture, salary, funding, red flags
TAILOR	Claude Sonnet	Resume tailoring + cover letter writing per JD
GUARDIAN	No AI (pure JS)	Safety layer — budget, rate limits, blacklist, human queue
SUBMIT	Claude Haiku	Playwright-based form fill and submission (off by default)
LEDGER	No AI (pure JS)	Pipeline tracker — status, stats, follow-up reminders
COACH	Claude Sonnet	Interview prep — activated on phone_screen/interview stage change

---

Prerequisites

Before you start, install:

• Node.js v18 or later — nodejs.org
• pnpm — install with npm install -g pnpm after Node.js
• Git — git-scm.com

---

API Keys

OrbitApply requires two API keys.

1. Anthropic Claude (required)

All AI writing — resume tailoring, cover letters, company intelligence, interview coaching — runs on Claude.

1. Sign up at console.anthropic.com
2. Go to API Keys → Create Key
3. Copy the key — it starts with sk-ant-

2. Tavily (required)

SCOUT and RECON use Tavily for all web searches.

1. Sign up at tavily.com
2. Go to your dashboard → copy your API key
3. The key starts with tvly-

---

Installation

1. Clone the repository

git clone https://github.com/Orbitumaiopensource/Orbitapply.git
cd Orbitapply

2. Install dependencies

pnpm install

3. Set up your environment file

cp .env.example .env

Open .env and fill in your real keys:

ANTHROPIC_API_KEY=sk-ant-your-key-here
TAVILY_API_KEY=tvly-your-key-here
PORT=3000
NODE_ENV=development

Never commit .env to version control. It is already in .gitignore.

4. Set up your candidate profile

memory/profile.json is your personal data file. It is gitignored and must be created manually.

# Mac/Linux
mkdir -p memory && touch memory/profile.json

# Windows (PowerShell)
New-Item -ItemType Directory -Force -Path memory
New-Item -ItemType File -Path memory\profile.json

Paste this template into memory/profile.json and fill in your information:

{
  "name": "Your Full Name",
  "title": "Your Current Job Title",
  "email": "your@email.com",
  "phone": "+1 555 000 0000",
  "location": "City, State",
  "linkedinUrl": "https://linkedin.com/in/yourprofile",
  "githubUrl": "https://github.com/yourusername",
  "targetRoles": ["Director of AI", "Head of AI", "VP of AI"],
  "targetLocations": ["Remote", "New York, NY"],
  "remotePreference": "remote",
  "targetIndustries": ["SaaS", "Fintech", "Enterprise Software"],
  "salaryCurrency": "USD",
  "salaryMin": 150000,
  "salaryMax": 250000,
  "yearsExperience": 8,
  "skills": ["AI Strategy", "LLM", "Agentic AI", "Python", "Node.js", "AWS"],
  "coverLetterTone": "professional",
  "education": [
    {
      "degree": "B.S. Computer Science",
      "school": "University Name",
      "year": 2016
    }
  ],
  "resume": {
    "path": "memory/resume.md",
    "summary": "One to two sentence professional summary."
  }
}

5. Create your base resume

Create memory/resume.md with your resume in Markdown format. TAILOR rewrites bullet language for each job — it never invents experience, so keep this file factual and complete.

# Mac/Linux
touch memory/resume.md

# Windows (PowerShell)
New-Item -ItemType File -Path memory\resume.md

6. Create required runtime directories and seed files

These directories and files are gitignored and must be created locally before first run:

# Mac/Linux
mkdir -p sessions credentials logs Apply
echo '{}' > sessions/sessions.json
echo '{"companies":[]}' > memory/blacklist.json
echo '{"contacts":[]}' > memory/protected.json

# Windows (PowerShell)
foreach ($dir in @("sessions","credentials","logs","Apply")) {
  New-Item -ItemType Directory -Force -Path $dir
}
Set-Content -Path sessions\sessions.json -Value '{}'
Set-Content -Path memory\blacklist.json -Value '{"companies":[]}'
Set-Content -Path memory\protected.json -Value '{"contacts":[]}'

---

Running the app

Development mode (recommended — auto-restarts on file changes)

pnpm dev

Production mode

pnpm start

Open your browser:

http://localhost:3000

No login required. The app is bound to 127.0.0.1 and is never accessible from the network.

---

Daily usage

1. Open http://localhost:3000
2. Click Run and set your goal (role, location, urgency)
3. Watch the live progress: SCOUT → RECON → TAILOR → LEDGER
4. Review SCOUT results — approve or reject individual jobs if you wish
5. Check the Documents section to review your tailored resume and cover letter for each job
6. When a real response arrives, update the application status in the pipeline board
7. When a stage reaches Phone Screen or Interview, COACH automatically generates your prep pack

---

Enabling auto-submit

SUBMIT is disabled by default. To enable automated form filling, add the following to orbitapply.json:

"submit": {
  "autoSubmit": true
}

With auto-submit enabled, GUARDIAN still enforces all limits. Applications with sensitive fields (salary, diversity, custom essay, security clearance) are always routed to the human review queue, never submitted automatically.

---

Blacklisting companies

To prevent OrbitApply from ever finding, tailoring, or applying to a company, add it to memory/blacklist.json:

{
  "companies": ["CompanyName", "another-company.com"]
}

GUARDIAN checks this list at the SCOUT filter stage and again before every TAILOR and SUBMIT action.

---

Budget controls

The default daily spend cap is $5 USD across all Anthropic API calls. Configurable in orbitapply.json:

"budget": {
  "dailyLimitUSD": 5,
  "hardStop": true,
  "alertAtUSD": 3
}

Approximate cost per full run (SCOUT + RECON + TAILOR for 5 jobs): $0.40–$0.80

---

Project structure

OrbitApply/
├── agents/              # Agent identity and rules (SOUL.md per agent)
│   ├── orbi/
│   ├── scout/
│   ├── recon/
│   ├── tailor/
│   ├── guardian/
│   ├── submit/
│   ├── ledger/
│   └── coach/
├── src/
│   ├── routes/          # Express API routes (/api/v1/*)
│   ├── services/        # Agent service modules (one file per agent)
│   └── utils/           # Logger, constants, file store
├── ui/                  # Frontend (served by Express at /)
├── memory/              # profile.json, resume.md, blacklist.json (gitignored)
├── sessions/            # Active session state (gitignored)
├── credentials/         # Auth profiles (gitignored)
├── logs/                # Runtime logs (gitignored)
├── Apply/               # All generated application documents
│   └── applications/
│       └── {Company} — {Job Title}/
│           ├── {Title}_{Company}_Resume_{Name}.doc
│           ├── {Title}_{Company}_Coverletter_{Name}.doc
│           └── metadata.json
├── workspace-scout/     # Daily job search results (gitignored)
├── workspace-recon/     # Company intelligence profiles (gitignored)
├── workspace-tailor/    # Fallback document workspace (gitignored)
├── workspace-ledger/    # pipeline.json — application tracker (gitignored)
├── workspace-coach/     # Interview prep packs (gitignored)
├── .env.example
├── orbitapply.json      # Agent configuration, budget, guardian settings
├── index.js             # Express server entry point
└── package.json

---

Logs

File	Contents
logs/run.log	Full pipeline run logs
logs/app.log	Server errors
logs/agent-actions.log	Every agent delegation and action

---

Security

• Server binds to 127.0.0.1 only — never accessible from the network
• All API keys are in .env — never hardcoded, never committed
• No authentication required — physical access to the machine is the only gate
• Never expose this app beyond localhost

---

Troubleshooting

pnpm: command not found — Run npm install -g pnpm first.

Cannot find module on startup — Run pnpm install to restore dependencies.

No target roles found in profile — Open memory/profile.json and make sure targetRoles is a non-empty array.

SCOUT returns zero results — Check your TAVILY_API_KEY in .env. Verify remaining credits on your Tavily dashboard.

Budget hard stop at startup — The $5 daily cap was already reached today. Either wait until midnight or raise dailyLimitUSD in orbitapply.json.

TAILOR documents look generic — Ensure memory/resume.md has complete, detailed content. Also check that your profile JSON accurately reflects your seniority and specialisation.

SUBMIT fails on a form — Disable autoSubmit for that session and submit manually. Review the screenshots in workspace-submit/ to see where the form failed.

---

License

OrbitumAI Free License — free to use, modify, and distribute. See LICENSE for details.

---

About the builder

Shuv Chowdhury — Founder, OrbitumAI

20+ years in enterprise transformation. 2,500+ hours in AI over the last two years. Now building AI products for non-technical founders and anyone who needs them.

• shuvchowdhury.com
• OrbitumAI
• LinkedIn

---

If it helps you land something — I'd genuinely love to hear about it.