How to Build an AI Job Search Agent With Claude Code in 2026 (The 29K-Star Open-Source Framework, Explained)

Verdict: The fastest way to run an autonomous AI job search in 2026 is to fork an open-source framework like ai-job-search (29,262 GitHub stars, MIT-licensed), feed Claude Code your CV and career goals, and let a drafter-reviewer agent pipeline evaluate postings you'd be a real fit for — then draft, critique, and revise your CV and cover letter before you hit submit. The framework's creator used exactly this loop to send 69 tailored applications, land 21 first interviews (~30% response rate), and start as an AI engineer in June 2026. The whole system runs on your laptop for the cost of a Claude Pro subscription ($20/month), and the human-in-the-loop submit step is what keeps it honest.

Last verified: 2026-08-04 · Framework: ai-job-search v.latest (MIT, 29,262 stars / 9,835 forks, created March 18, 2026) · Cost to run: $20/month Claude Pro minimum · Outcome (creator): AI engineer role, June 2026, after 69 applications / 21 first interviews.

TL;DR:

• ✅ The repository MadsLorentzen/ai-job-search is a Claude Code-native job application framework — fork it, fill in your profile, run three commands (/setup, /scrape, /apply).
• ✅ Its core insight is the drafter-reviewer pattern: one agent writes your CV/cover letter, a second agent reads it as a skeptical recruiter and flags hallucinations and missed keywords before you see it.
• ✅ The framework never fabricates skills you don't have — an explicit honesty constraint is baked into the prompt scaffolding.
• ✅ The human always clicks submit. The workflow deliberately stops before the apply button.
• ⚠️ Pricing and feature limits change — re-check the GitHub README and claude.com/pricing monthly.

---

Why "AI job search automation" is suddenly a real category

Job hunting in 2026 is a brutal numbers game with a quality floor. The structural problem is well-documented:

• 88% of employers admit their applicant tracking systems (ATS) filter out qualified candidates who don't match job descriptions exactly — the finding of a 2021 Harvard Business School / Accenture study of 8,720 "hidden workers" and 2,275 executives across the U.S., U.K., and Germany. (HBS Hidden Workers report)
• The same research put the population of "hidden workers" — people actively seeking work, qualified, and systematically filtered out — at 27 million Americans.
• A HireThing compilation of 2025 job-application statistics found that job seekers now submit 32 to 200+ applications on average before receiving an offer, with average time-to-hire stretching to 42 days and beyond. (HireThing 2025 job application statistics)

So three forces collide: ATS reject qualified people for keyword mismatches, candidates must apply at scale to get one response, and each tailored application takes 30+ minutes when done by hand. That's the exact problem an agent-based job search solves — not by spamming more applications, but by making each one honest, tailored, and submitted only when it actually fits you.

What is an AI job search agent?

An AI job search agent is a structured workflow (usually built on a coding-agent CLI like Claude Code, Codex, or Gemini CLI) that automates the four hardest parts of a job hunt: self-profiling, posting discovery and fit-scoring, tailored application drafting, and interview preparation. It is not a single chatbot prompt — it is a multi-agent system where specialized agents pass structured context to each other.

The key difference from "paste your CV into ChatGPT": the agent has persistent memory of who you are (your profile lives in files inside the repo), access to live job postings (via portal-scraping skills), and a review loop that catches the hallucinations and tonal flatness that ruin one-shot LLM drafts.

The ai-job-search framework is the most-starred example of this pattern. As of August 4, 2026 it has 29,262 stars and 9,835 forks on GitHub — a fork-to-star ratio of roughly 1:3, which signals that people are actually using it, not just bookmarking it. (GitHub: MadsLorentzen/ai-job-search)

How does the drafter-reviewer agent architecture actually work?

The architectural insight that makes the framework work is the drafter-reviewer pattern: one agent writes the application materials, a second agent with fresh context critiques them as a recruiter would, and the drafter revises. This is the same adversarial-review idea behind RAG "LLM-as-a-judge" systems — but applied to your job application instead of retrieval results.

Here's the loop, step by step:

1. Drafter agent reads your structured profile + the job posting, then drafts a tailored CV and cover letter (compiled to PDF via LaTeX — lualatex for the CV, xelatex for the cover letter).
2. Reviewer agent (separate context window) reads the draft as a skeptical recruiter: it checks for missed ATS keywords, flags any claim not supported by your real profile, calls out inflated experience, and verifies date consistency.
3. Drafter revises based on the critique. The framework imposes a hard honesty constraint — the AI cannot invent skills you don't have or inflate experience you haven't gained. If it can't honestly bridge a gap, it surfaces the gap as a gap and flags it for you.
4. ATS parseability check: if pdftotext (from poppler) is installed, the framework runs your draft through it to simulate how an ATS would parse the text. If poppler is missing, it degrades gracefully to a visual keyword review.
5. Human review and manual submit: the workflow deliberately stops here. You read every line, adjust tone, and you click the submit button yourself.

This five-step loop is why the framework's outputs survive adversarial scrutiny — and why the creator was transparent about using it with every employer, which often sparked technical conversations instead of counting against him. (LinkedIn: "I Automated My Job Search. It Made the Process More Human, Not Less." — Mads Lorentzen, PhD, March 23, 2026)

What are the 5 commands you actually run?

The user-facing surface of the framework is a small set of Claude Code slash commands. Each maps to a .agents/skills/<tool>/SKILL.md file — the same skill-package format Anthropic's Claude Code, OpenCode, and other agent CLIs now share. If you've ever read our guide to running an AI agent operating system for daily work, the pattern will look familiar: the skill files are the engine, the slash commands are the UX.

Command	What it does	When to run it
/setup	Builds your structured self-profile from a documents folder (CV PDF, LinkedIn export, diplomas, references), a single pasted CV, or an interactive interview. Idempotent — safe to re-run.	Once, at the start. The more documents you feed it, the better every downstream output.
/scrape	Searches job portals, deduplicates, and presents postings sorted by fit against your profile. Ships with Danish job-board skills (Jobindex, Jobnet, Akademikernes Jobbank) but /add-portal generates search skills for any local job board.	Daily or weekly, to surface new postings.
/rank	Batch-scores scraped postings against your fit framework when /scrape returns too many to eyeball. Returns a ranked shortlist.	When the scrape returns a flood.
/apply <url>	The full application pipeline: evaluate fit, draft CV + cover letter, reviewer critiques, drafter revises, compile to PDF, ATS check. You can also paste a job description directly if the URL is blocked.	One posting at a time — the human picks the match.
/interview	Generates a stage-specific interview prep pack and runs a mock interview using STAR-method examples pulled from your own profile.	After you land an interview.
/outcome	Records application results, archives materials, updates the tracker. /outcome followup surfaces quiet applications (default 10 days) and drafts follow-up messages.	After every interview, rejection, or offer.

Two optional commands round out the system: /expand enriches your profile by scanning linked public sources (GitHub, portfolio, Google Scholar) to surface competencies you might not think to mention, and /notion-sync / /gmail-sync pipe application status into Notion and Gmail for at-a-glance tracking.

How much does it cost to run an AI job search agent?

The framework itself is free and MIT-licensed. The only hard cost is the Claude Code subscription that powers the agents. As of August 2026, Anthropic's pricing is:

Plan	Price	What you get for job-search use
Free	$0	A small evaluation allowance — enough to try /setup, not enough for a real search.
Pro	$20/month ($17/month billed annually)	Full Claude Code CLI access; Sonnet 4.6 + Opus 4.7 + Haiku 4.5. The right tier for most job seekers — enough usage for daily scraping + a few /apply runs.
Max 5x	$100/month	~5x Pro's usage. For heavy daily scraping + bulk /rank across hundreds of postings.
Max 20x	$200/month	~20x Pro's usage. Priority during peak demand. Only worth it for very high-volume search sprints.
API	Pay-per-token	Sonnet 4.6 is ~$3 per million input tokens / $15 per million output tokens. Best for automation pipelines that run on a schedule.

Source: claude.com/pricing, confirmed via InventiveHQ's May 2026 Claude Code pricing breakdown. Pricing is volatile — confirm at the source before subscribing.

A practical note on cost: /apply plus the reviewer agent plus parallel /rank runs consume tokens fast, because each posting triggers a draft, a critique, and a revision. If you're on Pro, treat /scrape as your daily driver and save /apply for postings you'd genuinely send an application to. That alone keeps a job search under $20/month for most people.

How to set up the AI job search framework (step by step)

These are the prerequisites and the exact fork-to-first-application path, distilled from the repo README and the creator's own write-up.

Prerequisites

• Claude Code (CLI) — or a compatible agent CLI like Codex, Google Antigravity, or Gemini CLI via the repo's AGENTS.md. (See our guide to setting up an AI agent operating system: loops and plugging in new models for how these CLIs compare.)
• Python 3.10+ — for the portal-search CLI tools.
• Bun — the JS runtime the portal-search skills are written against.
• A LaTeX distribution with lualatex and xelatex (TeX Live, MacTeX, TinyTeX, or MiKTeX). The CV compiles with lualatex (pdflatex fails on modern MiKTeX with fontawesome5 errors); the cover letter compiles with xelatex (cover.cls requires fontspec).
• Optional: pdftotext from poppler — for the /apply ATS parseability check. Without it the framework degrades to a visual keyword review.

The fork-to-first-application path

1. Fork and clone. Fork first, don't just clone — the framework is designed around your personal profile data living in your fork. (GitHub: MadsLorentzen/ai-job-search)

   gh repo fork MadsLorentzen/ai-job-search --clone
   cd ai-job-search
   

2. Install the portal-search tooling. From the repo root:

   for tool in jobbank-search jobdanmark-search jobindex-search jobnet-search linkedin-search freehire-search; do
     (cd .agents/skills/$tool/cli && bun install)
   done
   

3. Run /setup. Drop your CV PDF, LinkedIn export, diplomas, and any reference letters into the documents/ folder, then inside Claude Code run /setup. Pick documents-folder mode. The more you invest here, the better every downstream output — the creator says the structured self-reflection in /setup delivered more value than the automated applications themselves.
4. Add your local job boards. If you're not in Denmark, run /add-portal with the job sites you actually use. The system investigates URL patterns, result structures, and access rules, then scaffolds CLI skills and test-runs queries. LinkedIn search works out of the box.
5. Run /scrape, pick a match, run /apply <url>. Read the draft the framework produces. Adjust tone. Then you open the portal and click submit.

A security note from the README: postings are treated as untrusted input — no embedded instructions inside a job description are followed, and no links inside the body are fetched. But the agent's defenses are instruction-level, not a sandbox. Always skim what was fetched on unfamiliar job boards before sending.

What was the actual outcome — does it work?

The honest, sourced answer: the framework's creator, Mads Lorentzen (a geophysicist whose industry position was cut in late 2025), used his own workflow from January through June 2026 and reports sixty-nine tailored applications, twenty-one first interviews, and one signed contract — he started as an AI engineer in June 2026. That's a roughly 30% application-to-first-interview conversion rate, against an industry baseline where most candidates are ghosted by default. (GitHub repo README, "Does It Actually Work?")

A nuance worth flagging: his earlier March 2026 LinkedIn article reported ~30 targeted applications, several interviews including a second-round, and explicitly no job offer yet at that point — the 69-applications / offer-secured outcome is the updated figure from the GitHub repo as of August 2026. The framework improved his results over time as he iterated on the drafter-reviewer loop. (LinkedIn, March 23, 2026)

The single most-quoted line from his writeup captures the philosophy: "The fit evaluation is worth more than any polished cover letter. The career path mapping is worth more than a hundred tailored CVs. The structured self-knowledge is the real product. The applications are just how it gets delivered."

How does this compare to other AI job-search tools?

The drafter-reviewer, honest-keyword, fit-first approach is genuinely different from the mass-apply tools that dominate the space. Here's the landscape, simplified:

Tool / repo	Pattern	Philosophy	Cost
MadsLorentzen/ai-job-search (29.3K★, MIT)	Drafter-reviewer agent pipeline, fit-scored, human submit	Fit over volume. Honesty constraint.	$20/mo Claude Pro
santifer/career-ops	Multi-CLI (Claude, Gemini, Codex, OpenCode), A–F scoring, application tracker	Same fit-first idea, broader CLI support	Free + your CLI's cost
Jobright Agent	Scans 400K+ roles daily, auto-tailors resume, one-click apply	Mass-market convenience	Freemium
ApplyPilot ("spray-and-pray")	Auto-applies to 1000+ jobs	Volume-first; HN comment: "how many interviews would you get applying to 1000 jobs?"	Free + Reddit-stunt history
Claude Cowork scheduled task	Daily 7am scrape + score + draft queue	Scheduled convenience on Claude Desktop	$20/mo Claude Pro

The honest distinction: tools built around volume versus tools built around fit. The 29K-star framework is firmly in the fit camp — and the market signal (its star count, the 1:3 fork-to-star ratio, the testimonials from job seekers worldwide) suggests that's where practitioners are moving. If you've read our breakdown of how to build a self-improving AI agent operating system, the pattern is the same: the agent's value is in the feedback loop (/outcome feeds back into /rank), not in raw throughput.

What this means for you

Three concrete takeaways depending on who you are:

If you're job hunting right now: Fork the repo, invest 30+ minutes in /setup (this is the step everyone underinvests in — the structured self-profile is the real product), run /scrape daily, and use /apply only on postings the fit evaluation greenlights. Keep the human submit step. The honesty constraint is the feature that keeps your applications credible — turn it off and you become another spray-and-pray bot that recruiters filter on sight.

If you're a small business hiring: The same drafter-reviewer pattern generalizes. If your team writes anything a human will judge — grant proposals, client pitches, contractor SOWs, RFP responses — a second agent with fresh context reading as a skeptical reviewer catches what a single-pass draft misses. The framework's .agents/skills/ directory is a working template for this pattern. Pair it with the free AI API providers we compared here to run the loop without piling up token costs.

If you're a builder: This repo is also a masterclass in the new "ship your skills" distribution model — you build a workflow for yourself, package it as a forkable repo of SKILL.md files (roughly 50% code, 50% plain-text instructions), and the open-source community amplifies it. The creator's other repos got dozens of stars; this one got 29,000. The difference was that he solved a real, painful problem for himself first and then gave the solution away. That's the same shift we wrote about in how to start a one-person AI consulting business: the moat in 2026 is the reusable, sharable workflow, not the model subscription.

Limitations and risks you should know

• LaTeX fragility. The CV template (moderncv + fontawesome5) breaks on the wrong LaTeX engine. The README mandates lualatex for the CV and xelatex for the cover letter — swapping engines will cost you an afternoon of debugging.
• Portal blocking. Some job-board URLs return 403 to scrapers. The framework lets you paste the job description directly into /apply as a workaround, but it's a manual step.
• LinkedIn Terms of Service. Automated LinkedIn search at volume is against LinkedIn's ToS. Use the LinkedIn skill sparingly, or rely on direct portal scrapers instead.
• Token consumption. /apply + reviewer + parallel /rank agents consume tokens fast. On the free Claude tier you'll exhaust your allowance in days; on Pro, budget for occasional overages during heavy weeks.
• No published releases. The repo has tags but no formal GitHub Releases — track main and re-pull periodically. There are usually a couple of open issues; check the Issues tab before depending on edge cases.
• It's a personal workflow framework, not HR-compliant software. You remain responsible for factual accuracy in anything you submit. The honesty constraint reduces hallucinations; it doesn't eliminate them. Read every line before you send.

FAQ

Q: Can an AI job search agent actually get me a job? A: It can dramatically improve the quality and fit of your applications — the framework's creator went from laid off to AI engineer in ~4 months on 69 applications with a ~30% interview-conversion rate. It is a force multiplier, not an autopilot: you still pick the postings, review every draft, and click submit yourself.

Q: Is the ai-job-search framework really free? A: Yes — it's MIT-licensed and free to fork and use. The only hard cost is a Claude Code subscription (from $20/month Pro) to power the agents. You can also run it against Codex, Gemini CLI, or OpenCode via the repo's AGENTS.md if you prefer a different agent CLI.

Q: Will the AI fabricate skills I don't have on my resume? A: No — an explicit honesty constraint is baked into the prompt scaffolding. The drafter agent cannot invent skills you don't have or inflate experience you haven't gained. If it can't honestly bridge a gap, it surfaces the gap and flags it for you. A separate reviewer agent also checks every draft for unsupported claims before you see it.

Q: Does this work outside Denmark? A: Yes for the core workflow (self-profiling, fit evaluation, drafter-reviewer pipeline) — that part is language- and country-agnostic. The portal-search skills ship with Danish job boards, but /add-portal generates search skills for any job site you point it at. LinkedIn search works out of the box.

Q: How is this different from just pasting my CV into ChatGPT? A: Three ways. (1) The framework has persistent memory of who you are — your profile lives in files, not a chat window you lose. (2) It has a drafter-reviewer loop that catches hallucinations and tonal flatness before you see the draft. (3) It has a fit-evaluation step that answers "should you even apply for this job?" before drafting anything — which the creator says was the most valuable part of the whole system.

Q: Is using AI to apply for jobs ethical? A: The framework's design answers this directly: it's designed to be transparent. The creator disclosed his use of it to every employer, which often became a technical talking point rather than a black mark. The honesty constraint means it reframes your real experience, it doesn't invent a fake candidate. The line is between "AI that helps you apply honestly" and "AI that lies for you" — this framework is built for the former.

Sources

• GitHub — MadsLorentzen/ai-job-search (29,262 stars / 9,835 forks, MIT license, created March 18, 2026): https://github.com/MadsLorentzen/ai-job-search
• LinkedIn — "I Automated My Job Search. It Made the Process More Human, Not Less." by Mads Lorentzen, PhD (March 23, 2026): https://www.linkedin.com/pulse/i-automated-my-job-search-made-process-more-human-mads-lorentzen-phd-38bje
• Harvard Business School / Accenture — "Hidden Workers: Untapped Talent" (2021, Figure 7 and Figure 9): https://www.hbs.edu/ris/Publication%20Files/hidden-workers09032021_Fuller_white_paper_33a2047f-41dd-47b1-9a8d-bd08cf3bfa94.pdf
• HireThing — "2025 Job Application Statistics — Updated Data You Need to Know for 2026": https://blog.hiringthing.com/2025-job-application-statistics-updated-data-you-need-to-know
• Anthropic — Claude plans & pricing (August 2026): https://claude.com/pricing
• InventiveHQ — "Claude Code Pricing 2026: Free, Pro, Max & API Cost" (May 30, 2026): https://inventivehq.com/blog/claude-code-pricing-explained
• AgentConn — "ai-job-search: GitHub's Hottest Agent Applies For You" (July 9, 2026): https://agentconn.com/blog/ai-job-search-agent-drafter-reviewer-claude-code-2026/

Updates & Corrections

• 2026-08-04 — Article first published. All stats (stars, forks, applications, interviews, pricing) verified against primary sources on August 4, 2026. Pricing/limits flagged as volatile; re-check claude.com/pricing and the GitHub README monthly.

---