← Projects

OfferFinch

A local macOS job-search assistant: finds matching jobs, scores the fit, and drafts a CV, letter and answers checked against your own profile.

OfferFinch is a job-search assistant that runs on one Mac. It searches job boards in the background, scores each posting against your profile, and when you press Prepare on a match, drafts a tailored CV, cover letter and application answers. Nothing is sent without your approval, and there are no accounts or hosted services: it is a menu-bar app with a local database.

Jobs come from job-board search (Greenhouse, Lever, Ashby, SmartRecruiters, Workable, Recruitee, Teamtailor and Personio postings, read from each board’s public data), LinkedIn, Indeed, Glassdoor, remote-job feeds and web search. Every job lands in the Inbox with a fit score and what matches or is missing from your profile.

How it works

From a scan to an approved application Scans run in the background. A match waits in the Inbox until you press Prepare; every draft is checked against your profile by a different vendor's model before you see it.

Parts: Scheduled scan (hourly or Scan now), Job boards (Greenhouse, Lever, LinkedIn…), Filter and fit score (hard rules, then a model), Inbox (matches wait here), Draft (CV, letter, answers), Claims check (a different vendor's model), Your profile (the only source of facts), Review and approve (nothing sent without you).

  • Scheduled scan to Job boards
  • Job boards to Filter and fit score
  • Filter and fit score to Inbox
  • Inbox to Draft (Prepare)
  • Draft to Claims check
  • Your profile to Claims check (quotes)
  • Claims check to Review and approve

The part I cared most about is the claims check. Your profile, imported from a resume and reviewed record by record, is the only thing a draft may claim. After a model writes the CV and letter, a model from a different vendor reads every line and has to tie each fact to a quote from the profile. A line it can’t back is flagged for you instead of being sent; if facts are missing, the app asks you a question rather than inventing an answer. Editing a line re-checks only the lines that changed.

Models run on AI plans you already have (Claude Code, Codex, OpenCode, and Gemini through Google’s Antigravity), never on API keys, and each step can use a different one.

Built to run unattended

The app is Electron with a React interface. A utility-process worker owns the SQLite database, the source readers and the model calls, and the interface reaches it through oRPC over a private MessagePort, so there is no HTTP server or open port. Work runs on a durable runner that stores runs, steps, leases and events in SQLite: after a crash, completed steps are not repeated and the job picks up where it stopped. Closing the window leaves scans running; Quit stops them.

Everything stays on the Mac, and Settings can export a full backup as one zip and restore it on another machine. A built-in demo, Try it with sample data, opens the app with a made-up profile and eight sample jobs scored and drafted by test models, with no AI connection or network. Every screenshot here comes from that demo.