PROJECT / 03 FEATURE ARTICLE

Job SearchAgent

TYPE AI SystemSTATUS active

A local-first Germany-focused job-search system with a macOS desktop app, Codex-assisted analysis, and SQLite tracking.

01 The operating sequence

Dashboard workflow

  1. 01Desktop app
  2. 02Discovery
  3. 03Human review
  4. 04Codex Desktop
  5. 05Materials
  6. 06Manual apply
  7. 07Tracking

02 Project overview

A desktop workflow with human checkpoints

The normal interface is the macOS Electron desktop application. Electron wraps and launches the existing Next.js Dashboard, which supports a Germany-focused job search and keeps discovery, review, preparation, and tracking in one local workflow.

Discovery is started manually. A person reviews each role, can ask Codex Desktop to analyze an interested role, and decides whether to prepare materials. Application submission remains manual.

03 How it works

From listing to follow-through

  1. 01

    Discover

    Manually run discovery across configured sources. Preference filters, normalization, canonical-source selection, and deduplication happen before accepted jobs enter local state.

  2. 02

    Review

    Each accepted role waits for a human decision: mark it Interested or Skip. Discovery does not score jobs automatically.

  3. 03

    Analyze

    For an Interested role, the Dashboard provides a job-specific prompt for Codex Desktop. The returned analysis is validated against the job description and references to the factual candidate profile before it is stored.

  4. 04

    Prepare

    Material generation is a deliberate action after validation. The project builds factual resume data from the approved tailoring plan; pinned RenderCV v2.8 renders the resume PDF.

  5. 05

    Apply manually

    The user opens the employer’s posting, submits the application there, and records application status in the Dashboard. The system never submits it automatically.

04 Implementation

A local desktop product layer

SQLite remains the canonical operational state for job records, source links, review and application status, and audit history. Local profile, configuration, and generated materials stay in the prepared private workspace.

Codex Desktop is an external analysis and tailoring orchestrator. There is no runtime LLM inside Electron or Next.js. Discovery stops at the review queue, and applications are submitted by the user outside the app.

The application includes a v0.9.1 Apple Silicon / arm64 DMG. It is unsigned, not notarized, and is not a clean-machine standalone installer: the prepared local workspace and Node project dependencies remain required. Python and the RenderCV environment are needed for resume generation; Bun is needed only for optional LinkedIn discovery.

05 Discovery sources

Configured paths into the review queue

  • Bundesagentur für Arbeit
  • Bounded web search
  • Optional LinkedIn public jobs
  • Bun required only for this path
  • Configured Greenhouse
  • Configured Lever
  • Configured Personio
  • Manual discovery
  • Human review
  • Manual application submission

There is no direct StepStone, Indeed, or XING integration. Source coverage depends on the configured search and company boards.

06 Media / output

A desktop workflow with explicit handoffs

Desktop app

Electron launches the existing Next.js Dashboard as the normal macOS interface.

Discovery

Manually run configured sources and send accepted, deduplicated listings to review.

Human review

Choose Interested or Skip before an external Codex Desktop analysis.

Materials

Generate from an approved tailoring plan with the local profile and RenderCV environment.

Manual apply

Submit on the employer's site, then track the application in local SQLite state.

Local state

Keep operational tracking, private configuration, and generated materials in the prepared workspace.

Discovery, analysis, preparation, and submission remain separate steps with explicit human decisions.