The Open-Source Job Hunt That Runs Inside Your AI Coding CLI

Every few months, another platform promises to automate the job search out of existence. Loopcv aggregates more than twenty job boards and applies on behalf of users through browser extensions and direct ATS form submissions, claiming users typically receive three times more interviews than manual applicants [1]. Careerflow.ai serves over 1.2 million job seekers with an all-in-one suite that includes AI resume building, application autofill, and LinkedIn optimization, reporting sixty percent faster time to interviews and twice as many offers [4]. Jobright advertises two million trusted users, eight million total jobs, and one-click application autofill across major ATS platforms [7]. JobCopilot markets itself as the world’s leading automation platform for job applications, promising ten times more interviews by applying to jobs across over half a million companies daily [10]. Kickresume, with a reported eight million users, trains its AI on thousands of real resumes to tailor CVs to specific advertisements [11]. The consensus across these services is simple: job searching is a volume problem, and the solution is to remove friction from the apply button.

Career-ops, an open-source project released in early 2026 by Santiago Fernández de Valderrama Aparicio, treats the premise as a category error. Fernández de Valderrama spent his own job search evaluating more than seven hundred forty listings, submitting sixty-eight applications, completing twelve interviews, and ultimately signing an offer for a Head of Applied AI role [6]. Rather than concluding that he should have applied to more jobs faster, he concluded that most listings were never worth his time. He built a system to find the few that were, then open-sourced it on April 4, 2026 [6]. The repository has since accumulated roughly seventy thousand GitHub stars, over thirteen thousand forks, and a Discord community exceeding four thousand members, with coverage in France, China, and Korea [3][6]. The hype is not for another auto-applier. It is for a tool that actively discourages applying.

The genuinely unusual technical idea is the runtime environment. Career-ops is not a web application, a browser extension, or a hosted dashboard. It is a collection of skill files, prompt templates, and configuration schemas that execute inside AI coding command-line interfaces such as Claude Code, Gemini CLI, OpenCode, and soon Codex [3][9]. A user pastes a job URL or description into a terminal session they were already using to write code, and the system triggers a local agentic pipeline: scrape the posting, classify its archetype, evaluate fit against the user’s CV, draft an ATS-optimized PDF resume, and append the result to a local Markdown tracker [9]. All personal data remains in Markdown tables, YAML files, and TSV batches on the user’s machine. There are no cloud accounts, no telemetry, and no subscription tiers [3].

This matters because it inverts the typical architecture of job-search software. Even open-source alternatives like ResumeLM present themselves as self-hostable web applications built on Next.js, React Server Components, PostgreSQL, and Supabase, with a Pro plan at twenty dollars per month for premium model access [2]. Career-ops skips the frontend entirely. It treats the AI coding CLI itself as the application runtime, leveraging the same context window, reasoning engine, and file-system access that developers already use for software engineering. The system is organized into twelve operational modes—each an isolated skill file with its own context, rules, and tools—for tasks ranging from single-offer evaluation and batch processing to portal scanning, LinkedIn outreach, and deep company research [6]. Because the agent reads and writes the same local files it consumes, users can instruct the CLI to modify its own scoring weights, archetypes, or negotiation scripts, and the model knows precisely which configuration to edit without a separate settings panel.

The scoring philosophy is the soul of the project. Career-ops evaluates listings across multiple weighted dimensions. Sources describe both a published five-dimension rubric—assessing match, north-star alignment, compensation, cultural signals, and red flags—and a more granular ten-dimension breakdown covering role fit, level fit, compensation band, growth trajectory, team signals, company stability, location and remote terms, mission alignment, tech stack overlap, and personalization room [3][9]. The model renders a holistic global score on a 1–5 or A–F scale through judgment rather than arithmetic, and it cites specific lines from the user’s CV against explicit job requirements [3]. Anything below 4.0 out of 5 is flagged as a hard pass. The documentation explicitly warns that the system is a filter, not a spray-and-pray tool, and that both candidate time and recruiter attention deserve conservation.

For the minority of listings that survive the filter, the pipeline generates a six-block evaluation report: role summary, CV match analysis, level strategy, compensation research drawn from sources such as Levels.fyi and Glassdoor, personalization hooks for the cover letter, and interview preparation framed through STAR+Reflection storytelling [9]. The system also accumulates five to ten master behavioral stories across evaluations, building a reusable story bank that answers any behavioral question without starting from scratch. A Go-based terminal dashboard built with Bubble Tea and Lipgloss offers a local TUI for browsing, filtering, and sorting the pipeline, though the primary interaction remains conversational inside the AI CLI.

The portal scanner reinforces the local-first, low-volume ethos. Pre-configured scrapers target dozens of company career pages and major applicant tracking systems including Greenhouse, Ashby, Lever, and Wellfound [9]. A verification pass using Playwright can check for stale postings that linger in public APIs after roles have closed, dropping expired entries before they contaminate the pipeline. The PDF generator also uses Playwright to render tailored resumes from an HTML template, injecting keywords for ATS compatibility while using only information already present in the user’s CV. The result is a document customized to the role, not fabricated for it.

Career-ops is unapologetically rough around the edges. The documentation admits that the first evaluations are rarely useful because the system begins as a stranger to the user’s career; it must be fed context, proof points, preferences, and aversions before its judgment improves. The legal disclaimer warns that AI models may hallucinate skills, that users bear responsibility for complying with third-party terms of service, and that the tool is provided as-is without warranty. There is no glossy onboarding, no mobile application, and no customer support chat. It assumes the user already lives in a terminal, maintains a Markdown CV, and is comfortable letting a headless browser navigate career pages under agentic control.

In the broader landscape, the project occupies a narrow but telling niche. It is not a resume builder like Kickresume, an autofill utility like Jobright, or an application bot like Loopcv [1][7][11]. It is closer to a personal operating system for job searching, implemented as prompts and configuration files rather than as a commercial SaaS wrapper. Its impact lies in demonstrating that agentic workflows do not need cloud hosting, user accounts, or venture-funded growth loops to be effective. By keeping data local and judgment human-in-the-loop, it offers a deliberate counterweight to the prevailing automation trend that treats recruiting as a numbers game to be won by brute-force submission.

The project remains permanently free and community-funded through GitHub Sponsors, with explicit guardrails against auto-submission to protect both candidate reputation and recruiter inboxes [3]. Planned additions include deeper integration with Codex and continued refinement of the CLI-native workflow. Whether it becomes a standard toolchain for technical job seekers or remains a bespoke artifact of one engineer’s successful hunt, it has already validated its central thesis. Sometimes the most effective way to win a job search is to engineer a system that tells you when to walk away.