# The 66,000-Star Job-Search Agent That Refuses to Auto-Apply > Source: > Published: 2026-08-21 00:08:28+00:00 [AI](https://sourcefeed.dev/c/ai)Article # The 66,000-Star Job-Search Agent That Refuses to Auto-Apply Career-ops runs inside your coding CLI and bets that ruthless filtering beats spraying applications. [Mariana Souza](https://sourcefeed.dev/u/mariana_souza) Every few months a job-search automation tool goes viral on GitHub, and the pitch is usually the same: apply to a thousand jobs while you sleep. [Career-ops](https://github.com/santifer/career-ops), which has piled up 66,000 stars since its April launch, is interesting because it makes the opposite bet. It's an AI job-search pipeline that refuses to submit anything on your behalf — and its README spends real estate telling you *not* to apply to most of what it finds. The other reason it's worth your attention has little to do with job hunting. Career-ops isn't an app. It's a repo of markdown "modes," YAML config, and Node scripts that you drop into whatever AI coding CLI you already run — [Claude Code](https://claude.com/claude-code), Codex, OpenCode, Antigravity, Qwen, Grok's CLI. That distribution model is quietly becoming a category, and career-ops is the clearest evidence yet that it works at scale. ## A filter, not a cannon The design center is a structured evaluation: blocks A through F score a listing across five weighted dimensions — role fit against your actual CV, level strategy, comp research, and so on — into a 1.0–5.0 score. A separate block G assesses posting legitimacy (scams, ghost jobs, suspicious reposts) without touching the numeric score, and an explicit no-sponsorship line in a JD gets flagged as a hard blocker. The system's stated policy: don't apply below 4.0 out of 5. Creator Santiago Fernández de Valderrama Aparicio built it during his own search and published the funnel: 740 listings evaluated, 68 applications sent, 12 interview processes, one signed offer — a Head of Applied AI role. The story got picked up by Business Insider and WIRED's Greek edition, which explains some of the star velocity. One person's funnel isn't a benchmark, but the ratio is the argument: the tool exists to collapse 740 into 68, not to inflate 68 into 740. That's a direct repudiation of the previous generation. Auto-apply tools — AIHawk went viral in 2024 doing exactly this for LinkedIn — fed an arms race where candidates spray AI-written applications and companies deploy AI screeners to cope, degrading the signal for everyone. Career-ops drafts cover letters, application emails, and even LinkedIn outreach messages, but every one dead-ends at a human approval gate. It never sends, submits, or clicks. Whether that's principle or liability management, it's the right call: the spray-and-pray tools are why hiring pipelines are drowning, and a filter only delivers value while it stays a filter. ## What running it actually looks like Adoption is one command — `npx @santifer/career-ops init` — which scaffolds a project directory. You `cd` in, open your coding CLI, and the system onboards you conversationally: CV, target roles, preferences. From there you paste a job URL to trigger the full pipeline (evaluation, tailored ATS-friendly PDF via [Playwright](https://playwright.dev), tracker entry), or run `/career-ops scan` to sweep portals. Batch mode fans out headless workers with `claude -p` or `opencode run` to evaluate ten-plus listings in parallel. A Go/Bubble Tea TUI gives you a dashboard over the pipeline, and analysis scripts surface rejection patterns and per-ATS advance rates from your own data. One architectural choice deserves credit: instead of scraping LinkedIn (fragile, ToS-hostile), the scanner leans on the ATS layer — [Greenhouse](https://www.greenhouse.com), Ashby, Lever, Wellfound — where job data is served from stable, public endpoints, with 55-plus provider modules and 100-plus companies preconfigured. That's the difference between a tool that breaks monthly and one that's shipped 20 npm releases since April and was pushed to again today. The trade-offs are real, though. Deep-evaluating hundreds of listings through a frontier model isn't free; the docs acknowledge this with a whole running-on-a-budget guide covering OpenRouter free tiers, [Ollama](https://ollama.com), and any OpenAI-compatible endpoint. Privacy cuts the same way: your CV, salary expectations, and interview history flow through whatever model API you've wired up — local-first architecture, but only as private as your model endpoint. And the quality of block-A-through-F judgments is bounded by how much context you feed it; the README is refreshingly honest that "the first evaluations won't be great." ## The real story is the runtime Here's the part that outlasts the job-search niche. Career-ops has no backend, no auth, no SaaS tier. Its "runtime" is your coding agent: the markdown modes are prompts, the scripts are tools, Playwright is the effector, and Claude Code (or any agent-skill-compatible CLI) supplies the reasoning loop, the sub-agent orchestration, and the billing relationship. The maintainer wrote a domain expert system; Anthropic, OpenAI, and Google operate the infrastructure. Sixty-six thousand stars says developers will happily adopt serious software shipped this way. Expect this shape everywhere: agent-native skill packs for personal finance, immigration paperwork, freelance ops — verticals where the logic is judgment-heavy, the data is personal, and a hosted service would be a privacy and liability nightmare. The losers are the VC-backed "AI job copilot" SaaS products, which now compete with a free MIT-licensed tool that's more capable and keeps your data on your disk. The catch is who gets to use it. Career-ops assumes a terminal, Node, and a paid coding-CLI subscription (or the patience to wire up local models) — which means the candidates best equipped to out-filter AI screeners are the ones who least need help. And its core advantage is self-limiting: the 4.0-threshold discipline works because most applicants haven't got it. If tools like this become as common as their auto-apply predecessors, hiring signal degrades again and the arms race resumes one level up. For now, though, if you're a developer job-hunting in this market, this is the rare viral repo where the hype and the utility actually line up. Just read the drafts before you hit send — that gate is the whole point. ## Sources & further reading - [santifer/career-ops](https://github.com/santifer/career-ops)— github.com - [career-ops: Open-source AI job search agent, local-first](https://career-ops.org/)— career-ops.org - [career-ops: How I Built My Own AI Job Search Tool](https://santifer.io/career-ops-system)— santifer.io - [How I built a tool to filter job listings and landed a Head of AI role](https://www.businessinsider.com/how-i-built-tool-filter-job-listings-landed-head-ai-2026-4)— businessinsider.com - [@santifer/career-ops](https://www.npmjs.com/package/@santifer/career-ops)— npmjs.com [Mariana Souza](https://sourcefeed.dev/u/mariana_souza)· Senior Editor Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon. ## Discussion 0 No comments yet Be the first to weigh in.