What Is CareerRat?

The problem

Keyword-match job sourcing produces noise. Spray-and-pray applying produces silence. Most job-search tooling optimizes the wrong thing: volume over signal.

CareerRat is built around a different rule: do not tailor or apply until the job has passed a real body-read gate. Every skill in the system enforces that principle.

What CareerRat actually is

CareerRat is an agent runtime, a local workspace that a supported AI CLI runs inside. The agent reads a contract file (AGENTS.md) that tells it how to route your intent to the installed skills. Each skill is a detailed, executable procedure. You talk to the agent; the agent drives the skills. CareerRat owns the durable workflows and threads, while direct and ACP adapters let the selected local engine execute them.

The CLI (careerrat) scaffolds the workspace, serves the local app, checks health, and exposes validated config helpers. It does not do job-search work. The agent does.

The intended loop

1. Onboard — ingest-profile interviews you and saves private candidate state: targets, comp floor, keep/cut signals, evidence bank, honesty boundaries, and writing-style calibration. SQLite is authoritative in new workspaces; .careerrat/candidate/ files are compatibility exports.
2. Source — setup-searches + search-jobs build searches from your targets, combine built-in public job-board sources with AI open-web discovery, find additional specialist boards and employer pages, dedupe, and produce a coarse-triage queue. Every role family uses the same candidate rules. A credible result whose full description is not yet readable is saved as an unverified lead, not treated as a verified posting.
3. Gate — evaluate-job reads the full posting and returns a verdict: keep or cut the job, how well it fits, whether the pay clears your floor, and what to do next. Those four print as a GATE / FIT / COMP / ACTION block, joined by extra lines when they apply — a comp anchor to negotiate against, a comp estimate when no band was posted, a legitimacy warning on a suspect listing. Nothing gets tailored before a KEEP.
4. Tailor — tailor-application builds honest résumé, cover letter, and short-answer artifacts from your evidence bank only.
5. Communicate — email-comms drafts and tracks recruiter threads, follow-ups, scheduling, and negotiation.
6. Track — track-outcomes records results and triggers reevaluate-strategy when outcome thresholds trip.
7. Prep — interview-prep builds audience-segmented packets, coaches live negotiation, and captures debrief lessons.

The loop compounds: learnings from interview debriefs and rejections feed back into evaluation and tailoring so each pass is sharper than the last.

Longer searches and intake operations are owned by CareerRat rather than the current page. They keep running when you move between views, restore their durable status after reload, and surface a retry if the app restarts before the work finishes.

What "local by default" means

• SQLite state, compatibility exports, and artifacts are stored locally and excluded from version control by CareerRat.
• Relevant context sent through your selected AI runtime follows that provider's privacy and retention terms.
• No CareerRat-operated server receives your data.
• Private-account browser automation is opt-in per capability, uses your own browser session, and stores no credentials. Saved job-site search uses a site-specific Yes/No choice when a source is added or first used and login is needed, not the private-account permission matrix.

What CareerRat is not

• Not a resume builder (it tailors from your evidence; it does not write from scratch).
• Not a job board (it sources from boards you configure; it is not a board itself).
• Not a one-click mass-apply tool (the gate and confirm-first submission are intentional guardrails).
• Not a subscription service (MIT-licensed and free; install and run it locally).