jev-clerk
macOS bookkeeping clerk booking supplier invoices: Jev picks each click from a closed list; a deep model rewrites the playbook.
Overview
jev-clerk is a bookkeeping clerk that works a real Mac desktop: it opens a supplier invoice PDF in Preview, reads it, types it into the open-source desktop accounting app Frappe Books (which has no API), saves, submits and files the PDF. It uses two models at two speeds. Jev makes every decision without writing text: each step it receives the screen as OCR lines and answers two multiple-choice questions, which kind of action from a closed list of 12 and which line to click, at about 0.35 s and $0.0002 per decision.
A deep model (Claude Fable 5.1 in the run) never touches the screen; after every 10 invoices it reads the full decision log and rewrites the playbook: the wording of Jev's questions, lessons shown to Jev, typing formats, waits and reflexes that code runs without a model call. Code validates every rewrite so the action list stays closed. The referee is the accounting database, opened read-only. In a recorded run of 30 real invoices worked three times, early playbooks booked none correctly, while later versions reached 8/9, 9/9 and 6/8 correct USD invoices, with zero wasted steps at about 30-44 s per invoice.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which kind of action comes next? | — |
| Choice | Which OCR line should be clicked? | — |
How it works
- OCR the screen into lines
- Jev picks action type and target line
- Code executes clicks and typing
- Every 10 invoices, a deep model rewrites the playbook
- Database check decides whether an invoice counts
Strengths
- Cheap, fast per-step decisions
- Self-improving playbook with validated rewrites
- Objective database referee
Caveats
- Performance fluctuated across playbook versions
- macOS and Frappe Books specific
At a glance
- Type
- Service
- Status
- prototype
- Language
- Python
- Runtime
- macOS desktop automation
- Has code
- yes
Use cases and integrations
How curated lists describe it
A macOS desktop clerk that books supplier invoices: Jev picks each click from a closed action list; a deep model only rewrites the playbook.
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev decide?
The action type from a closed list of 12 and which screen line to act on.
What does the deep model do?
It rewrites the playbook every 10 invoices without touching the screen.
How is success measured?
By checking the accounting database for a correct submitted invoice.
Related projects
Loopfour Studio features rebuilt with Jev: code finds candidates, Jev picks with confidence, code assembles verbatim results.
ProjectprototypeTypeScriptRust streaming event bus whose routing is decided by Jev, one question per subscription, with deliver, review or drop thresholds.
ProjectbetaRustReusable project context and specialist workflows for existing AI tools, with an addon that offloads bounded decisions to Jev.
ProjectproductionTypeScriptVerifiable coding-agent runtime built on checkable execution graphs, listed with an optional Jev provider for bounded choices like routing.
ServiceproductionTypeScriptAutoGPT agent platform, listed as adding Jev decision blocks for typed routing, filtering, scoring and confidence-gated next actions.
ProjectproductionPythonRust MCP server for desktop control via accessibility trees, with an optional decision model (Jev or compatible) for small choices.
MCP serverproductionRust