Enkidu for macOS — verified Codex usage tracking and work planning

[yun-hanmin]

I kept running into a simple problem: Codex could show me a percentage, but that number alone did not tell me whether I had enough capacity for the work I was about to start.

So I built Enkidu, a native macOS companion for people who use Codex regularly.

The goal is not to create another percentage meter. It is to help answer a more practical question:

Given what I have left, should I start this task now, reduce its scope, use a more efficient approach, or wait for the reset?

Enkidu keeps verified Codex usage separate from the work plan you choose. It does not claim that a percentage, model recommendation, or pace estimate can guarantee that a task will finish.

*Verified capacity is kept separate from the user’s task plan. Enkidu does not inspect prompts or project content.

[











Feature overview]

Enkidu at a glance

Enkidu is a privacy-first native macOS Codex usage tracker and work-planning companion. It combines verified weekly usage remaining, reset timing, freshness, Recent and Typical Pace, local history and reports, account-available Model Advisor guidance, bilingual Value Max prompt presets, a menu-bar status view, and native macOS widgets.

• Platform: macOS, with a universal preview build for Apple Silicon and Intel Macs
• Languages: English and 한국어
• Usage data: Verified Codex allowance readings, reset timing, freshness, pace, and local history
• Planning: Work Profiles, user-selected Task Budget, Model Advisor, and Value Max prompt presets
• Privacy: Local normalized data only; no prompts, chats, project files, browser data, or credentials
• Connection: The user’s signed-in local Codex installation; no separate OpenAI API key required
• Current preview: Enkidu v0.9A, Build 35
• Official project page: https://www.hanminyun.com/enkidu

1. Insights

Insights turns the current Codex allowance into a practical capacity overview.

It brings together:

• verified weekly usage remaining;
• reset timing and last-sync freshness;
• recent usage pace;
• longer-term typical pace;
• whether the current pace appears sustainable until reset;
• account token activity when Codex provides it;
• clear current, saved, stale, and unavailable states.

Recent Pace and Typical Pace remain separate so that a heavy session is not hidden by a longer average. If there is not enough fresh verified history for a responsible estimate, Enkidu says that it is still learning instead of inventing a result.

2. Codex Integration

Enkidu communicates locally with the user’s signed-in Codex installation through the Codex App Server and an optional local Codex plugin.

The integration can provide:

• read-only usage and reset information;
• the selectable model catalog available to the current Codex account;
• privacy-limited lifecycle signals such as working, waiting, turn finished, or closed;
• optional prompt-aware model guidance based on a generic work profile;
• explicit task actions when the user chooses to invoke them.

Enkidu does not inspect the screen, scrape webpages, read browser cookies, collect credentials, or copy prompts and project files. A finished Codex turn is not presented as a completed project, and Enkidu does not automatically change Codex settings or models.

3. Model Advisor

Model Advisor helps users choose an appropriate currently available model and supported thinking level for the kind of work they plan to do.

Users can select a general work profile such as:

• Quick Fix;
• Everyday Coding;
• Hard Debugging;
• Visual Check;
• Long Build;
• Important Design Decision;
• Parallel Research;
• Save Capacity.

Enkidu compares that profile with verified remaining capacity, the user’s chosen task budget, and the current account-filtered Codex model catalog. It recommends one available model and supported thinking level, explains the reason in plain language, and shows a lower-usage alternative when one is available.

It never switches the active model automatically, recommends an unavailable thinking level, or promises an exact percentage of savings.

4. Value Max prompt presets

Value Max provides reusable prompt structures for both Codex and ChatGPT in English and Korean.

The available presets cover:

• Quick Fix;
• Build a Feature;
• Hard Debugging;
• Visual Check;
• Important Decision;
• Long Build / Release.

Users choose the destination, work type, prompt language, answer language, and level of detail. Every visible choice changes the prompt shown in the preview, and Copy Prompt copies that same text for use in the current conversation.

The presets remain deliberately prompt-blind. Enkidu stores only generic selections and preset identifiers. It does not receive or save the completed prompt, private project content, URLs, or conversation text.

5. Reports

Reports summarizes how usage, guidance, and confirmed work decisions develop across the current Codex reset cycle.

It can show:

• verified remaining allowance;
• measured pace and reading count;
• reset timing;
• a verified-history chart;
• privacy-safe daily activity;
• guidance delivered;
• guidance the user confirmed following;
• measured comparisons when enough comparable tasks exist;
• a conservative estimated-usage-managed result when the evidence threshold is met.

Suggestions, confirmed decisions, measured comparisons, and estimated results remain separate. Enkidu does not present an unconfirmed recommendation as a saving, and it does not claim that purchased credits were charged or saved.

Report data stays locally on the Mac and excludes prompts, task names, project paths, transcripts, credentials, cookies, and account identifiers.

6. Verified history

History is built only from periodic readings saved after successful usage checks.

Users can:

• review verified usage across selectable windows from 30 minutes to 7 days;
• inspect the verified time and remaining percentage of individual graph points;
• see reset boundaries and changes in pace;
• retain the last verified value safely when a new check fails;
• distinguish a current reading from a saved or stale one.

Changing the graph range changes only the visible window. It does not increase the sync frequency, interpolate missing values, or create artificial data. Enkidu does not present its history as a second-by-second live feed.

7. Privacy boundary

Enkidu is local and privacy-conscious by design.

It does not read, save, or transmit:

• prompts or chats;
• project files or source content;
• browser data or cookies;
• passwords, tokens, or credentials;
• email addresses or account identifiers;
• private task names or file paths.

It stores only the normalized local information needed for verified usage history, planning preferences, optional reports, widgets, and alerts.

What changed in Build 35

Build 35 improves reliability after an update, particularly synchronization between the main app and its macOS widgets.

The update prevents an older WidgetKit process from leaving placed widgets on a previous accepted reading after an in-place app replacement. It also reduces redundant widget reload requests and requests a refresh only after the final verified usage history and pace state are ready.

Try Enkidu

Download Enkidu Build 35:
https://www.hanminyun.com/enkidu

If you try it during a real Codex session, I would especially like to know:

Did Enkidu help you make a real decision—start the task, reduce its scope, choose a more efficient approach, or wait for the reset?

Enkidu is an independent project and is not affiliated with or endorsed by OpenAI.

Preview-build disclosure: Enkidu is currently an independently distributed macOS preview build. It is not available through the Mac App Store and has not been notarized by Apple.

You can email me at yuncinema@gmail.com for more in-depth discussions on how I've come up with this and where it's headed.

[yun-hanmin]

Frequently asked questions

Q1. What is Enkidu?

Enkidu is an independent native macOS companion for Codex users. It combines verified usage tracking, reset timing, pace guidance, local history, reports, model guidance, prompt presets, and work-planning controls.

Q2. Is Enkidu just a Codex percentage meter?

No. The verified percentage remains authoritative, but Enkidu combines it with freshness, recent and typical pace, a user-selected task budget, local history, reports, and conservative model guidance to support practical work decisions.

Q3. Does Enkidu read my prompts or project files?

No. Enkidu does not read or save prompts, chats, project files, browser data, cookies, credentials, email addresses, or account identifiers.

Q4. Does Enkidu require an OpenAI API key?

No separate API key is required. Enkidu communicates locally with the user’s signed-in Codex installation through the Codex App Server and optional local Enkidu integration components.

Q5. Does Enkidu automatically switch Codex models?

No. Model Advisor can recommend an account-available model and supported thinking level, but Enkidu never changes the active model automatically.

Q6. Is Enkidu’s usage history real-time?

No. History contains only periodic readings saved after successful checks. Enkidu does not invent missing points or present its graph as a second-by-second feed.

Q7. Can Enkidu guarantee that I have enough usage to finish a task?

No. Its pace estimates, budgets, and model recommendations are decision support. They do not guarantee task completion or exact usage savings.

Q8. Is Enkidu an official OpenAI product?

No. Enkidu is an independent project and is not affiliated with, endorsed by, certified by, or officially supported by OpenAI.