DEV Community: Michel Johnson

I got tired of sending half-written thoughts to cloud AI, so I built this for Mac

Michel Johnson — Mon, 05 Oct 2026 03:49:18 +0000

Every AI writing tool seemed to ask me to do something slightly ridiculous:

Write half an email.

Stop.

Copy it.

Open an AI tab.

Paste it.

Ask it to fix the sentence.

Copy the answer.

Go back.

Paste it.

Continue writing.

All of that because I wanted to fix a typo or make one sentence sound less awkward.

So I built DraftKey.

👉 https://www.getdraftkey.com/

The idea is pretty simple:

AI writing help should feel more like spellcheck and less like visiting another destination.

So DraftKey lives where I'm already typing

I'm writing an email and make a typo?

Keep typing.

I'm halfway through a sentence and blank on the next few words?

Hit Tab if the suggestion is good.

I wrote something that sounds way more aggressive than it did in my head?

⌘ + ⌥ → make it friendlier → review → insert.

No copy/paste pilgrimage required.

DraftKey currently gives you:

automatic spelling + grammar cleanup when you pause

AI autocomplete for the next few words

rewrites like shorter, clearer, friendlier, professional, etc.

custom rewrite instructions

per-app controls

and one completely unnecessary feature that I now refuse to remove…

The AI runs on your Mac

This was the part I cared about most.

Think about the text you type every day:

Client emails.

Internal messages.

PR descriptions.

Notes.

Unfinished ideas.

Things you would absolutely never publish.

It felt weird to me that fixing "teh" or rewriting a two-sentence email could require sending the entire thing to a cloud AI service.

So DraftKey's writing model runs locally on Apple Silicon.

Autocomplete, corrections, and rewrites are processed on the Mac.

After the model is downloaded, your actual writing doesn't need to be shipped off to a cloud AI API.

There are still internet-connected parts of the product — activation, updates, and downloading the model — because I'm not going to pretend the machine never touches the internet.

But the text you're actively writing?

That's the part I wanted to keep local.

The annoying engineering problem wasn't actually the AI

It was text fields.

If you've ever worked with macOS Accessibility APIs, you can probably guess where this went.

Not every app exposes text in exactly the same way.

Normal text fields can behave beautifully.

Custom editors can decide they have other plans for your afternoon.

Secure fields should obviously be left alone.

And sometimes an app technically exposes a text field while still doing something bizarre enough to make you question several life decisions.

So I'm intentionally not marketing DraftKey as:

WORKS PERFECTLY IN EVERY APP™ 🚀

It doesn't.

It works in compatible macOS text fields, and I'm continuing to improve compatibility as I find edge cases.

I would rather tell people where the limitation is than pretend macOS Accessibility is magic.

And then I made a terrible product-management decision

I added mechanical keyboard sounds.

Six of them.

Recorded key presses/releases.

Cream.

Topre.

Keychron Red.

Cherry MX Blue.

Cherry MX Black.

And more.

Then I added volume control.

Then tone control.

Then stereo positioning, so keys on the left side of your keyboard can sound slightly more to the left in your headphones.

Did my local AI writing assistant need spatially positioned mechanical keyboard sounds?

Absolutely not.

Was it worth it?

Unfortunately, yes.

You can actually try that part in the browser without buying or installing anything:

👉 https://www.getdraftkey.com/

Headphones recommended.

I apologize in advance if you lose 15 minutes comparing switches.

I also didn't want another AI subscription

DraftKey is currently sold as a lifetime license for up to three Macs.

No monthly meter running because you wrote too many emails this month.

The local model is about 1.12 GB, and you'll need:

Apple Silicon

macOS 14+

at least 8 GB RAM

Local AI obviously has tradeoffs.

It uses your hardware.

There's a model download.

Cloud models can be much larger.

But for the tiny writing tasks I do dozens of times every day, I've ended up really liking the trade:

smaller local model + almost zero workflow interruption + private text.

The thing I'm trying to build

I'm not trying to replace ChatGPT or Claude.

If I need to reason through something complicated, I want a full AI conversation.

DraftKey is for a different category of moment:

I already know what I want to say.
Just help me say it without breaking my flow.

That's the whole product thesis.

AI writing that feels less like opening an AI product...

…and more like another native capability of the text field.

If you're on an Apple Silicon Mac, I'd genuinely love for you to try it:

👉 https://www.getdraftkey.com/

And even if you have zero intention of buying it, go try the keyboard sounds on the site and tell me which switch wins.

I'm especially curious what other devs think about:

Local model vs. cloud model for something this ambient
Where you'd actually use autocomplete during the day
Which macOS apps you'd most want me to make sure work well

I'll be in the comments.

And yes, I am prepared to defend the completely unnecessary keyboard sounds.

I got tired of leaving what I was writing to use AI, so I built DraftKey

Michel Johnson — Thu, 01 Oct 2026 03:53:37 +0000

I’ve been working on a little Mac app called DraftKey, and I’m finally at the point where I want to share it with more people.

I built it because I kept running into the same annoying problem.

I’d be writing an email, message, document, or whatever else, and I’d either make a bunch of typos or get stuck trying to make a sentence sound right.

Then I’d open an AI chat, copy everything over, type something like “make this sound better,” copy the result, go back to what I was doing, paste it, fix a couple things, and continue.

It works, but it completely pulls you out of writing.

I wanted something that just helped while I was already typing.

So I made DraftKey.

https://www.getdraftkey.com/

What DraftKey actually does

DraftKey is a private AI writing assistant for Mac.

The basic idea is pretty simple.

You write normally, and DraftKey can clean up spelling and grammar when you pause.

If you’re halfway through a thought, it can suggest the next few words.

If a sentence sounds weird, you can select it and make it friendlier, shorter, clearer, more professional, or just tell DraftKey what you want changed.

Then you keep writing.

You don’t have to move everything into another editor or have a conversation with a chatbot every time you want help with one sentence.

That was really the whole reason I wanted to build it.

The AI runs on your Mac

This was probably the biggest thing for me.

A lot of what we type every day is stuff I don’t really want to send somewhere just to fix a typo.

Emails, work messages, random ideas, unfinished drafts, personal notes, things that make absolutely no sense without context.

So DraftKey’s writing model runs locally on your Mac.

Your text isn’t sent to a cloud AI service to generate corrections, autocomplete suggestions, or rewrites.

There are still a few things that need an internet connection, like downloading the model, activating your license, and checking for updates, but the actual writing AI runs on the device.

I really like knowing that I can use AI on something I’m writing without having to upload the writing itself.

I also gave it autocomplete

Autocomplete ended up being one of my favorite parts.

DraftKey can suggest what comes next while you’re typing.

If the suggestion is useful, hit Tab.

If it isn’t, just keep typing and ignore it.

I specifically didn’t want it to feel like AI was constantly trying to take over what I was writing.

It’s more like having a few words ready when you need them.

Sometimes it saves you two seconds.

Sometimes it somehow guesses exactly what you were about to say.

Either way, it’s become one of those things that feels weird when I turn it off.

Rewriting is there when you actually need it

There are also times when the sentence is technically fine, but you know it sounds bad.

Maybe an email sounds colder than you intended.

Maybe you used 40 words to say something that should have taken 15.

Maybe you need to turn a casual message into something you can send at work.

In DraftKey, you can use the text you already have and ask it to change the tone, length, or wording.

You can do things like:

Make this friendlier

Make this shorter

Make this clearer

Make this more professional

Fix the grammar

Or just type what you want it to change

You get to review the new version before anything gets inserted.

I wanted AI to help edit my writing, not quietly replace it.

And then I got carried away with keyboard sounds

This part has basically nothing to do with AI.

I added mechanical keyboard sounds.

There are six different recorded keyboard sound profiles, and you can change the volume, tone, and stereo positioning.

So keys on the left side of the keyboard can actually sound slightly more to the left in your headphones.

Did DraftKey need this feature?

No.

Do I now hate typing without it?

A little bit.

You can actually try the keyboard sounds on the website without downloading anything.

I wanted DraftKey to feel like a Mac app

Another thing I cared about was keeping the app simple.

I didn’t want another giant AI dashboard with 70 buttons, prompt libraries, agents, workspaces, projects, and a sidebar asking me to create a new chat.

DraftKey is supposed to sit quietly in the background and help when you need it.

That’s it.

Write.

Fix something.

Finish a thought.

Keep going.

Who I built it for

Honestly, I originally built it around the way I use my Mac.

But I think it makes sense for anyone who writes a lot during the day.

Developers writing documentation.

Founders sending an unreasonable number of emails.

Students.

Writers.

People working remotely.

Anyone who communicates through Slack, Discord, email, docs, or text all day.

And especially people who like using AI but would rather keep their actual writing on their own computer.

A couple technical details

DraftKey currently works on Apple Silicon Macs running macOS 14 or newer.

The AI features need at least 8 GB of memory.

The writing model is around 1.12 GB and gets downloaded during setup.

DraftKey works with compatible macOS text fields using Accessibility, so there are some editors where support depends on how that app handles its text fields. Secure fields are excluded.

There’s also a lifetime license for up to three Macs instead of another monthly AI subscription.

Anyway, that’s DraftKey

I’ve spent a lot of time building and tweaking this thing, and now I mostly want real people to use it and tell me what sucks.

If you’re on a Mac and any of this sounds useful, give it a try:

https://www.getdraftkey.com/

I’d especially love feedback on the autocomplete and automatic corrections, since those are the parts you really have to live with for a bit to know whether they feel right.

And if there’s something you wish it did, tell me.

There’s a pretty good chance I’ve either thought about building it already or I’m about to.

I asked AI: “what tools should every developer have in 2026?”

Michel Johnson — Thu, 16 Apr 2026 17:18:00 +0000

Here’s the answer it gave me — surprisingly practical.

AI coding assistant

Tools like GitHub Copilot → https://github.com/features/copilot
or Cursor IDE → https://cursor.com

Writes boilerplate, fixes bugs, explains code.
At this point it’s baseline.

GitHub Copilot alone is used directly inside your editor to suggest entire functions and workflows ()

Local + cloud LLM setup

Something like Ollama → https://github.com/ollama/ollama

APIs like OpenAI API → https://platform.openai.com

Local for speed/privacy, cloud for power.

A better terminal

Warp Terminal → https://warp.dev

Modern terminal built for AI workflows and automation, used by hundreds of thousands of developers ()

AI usage / cost tracking

This is where most devs get blindsided.

If you’re building with APIs, you need to see:

how many tokens you’re using
what each session is costing
where spikes are happening

Tools like TokenBar → https://tokenbar.site

It just sits in your menu bar and shows usage live.
You catch runaway prompts early instead of after the bill hits.

API testing workflow

Postman → https://postman.com
or Insomnia → https://insomnia.rest

Faster iteration, easier debugging.

Fast deployment

Vercel → https://vercel.com
/ Railway → https://railway.app

Ship instantly instead of overthinking infra.

Knowledge capture

Docs, notes, saved snippets.

You forget more than you think.

Focus tooling

Anything that blocks distractions while you work.

Direct impact on output.

Final thought

The stack is shifting from:
editor + framework + cloud

to:
AI + infra + visibility

Most people upgraded the first two.
Very few upgraded the last one.

Curious what others are using in 2026.

🧰 Mac Dev Stack: Tools I Actually Use Daily (2026)

Michel Johnson — Sun, 12 Apr 2026 17:32:24 +0000

I’ve tried a lot of Mac dev tools over the past year—most are either overhyped or redundant.

Here’s a tight list of tools that are actually worth using if you’re building daily.

⚡ Core workflow

Raycast
Replaces Spotlight completely.

Run scripts
Manage windows
AI commands
Extensions for everything

This has basically become my “OS inside the OS”.

💻 Coding

Cursor
AI-native editor that feels like what VS Code should have evolved into.

Edit entire codebases with prompts
Fast context awareness
Actually usable for real work
🖥️ Terminal

Warp
Modern terminal with:

Blocks instead of raw logs
Command suggestions
Better UX than iTerm
📦 Version control

GitHub Desktop
Simple but effective if you don’t want to live in CLI 24/7.

🧠 Knowledge / docs

Notion
Docs + notes + lightweight DBs.
I use it for:

product specs
quick notes
internal docs
📬 Communication

Slack
Still unavoidable for most teams.

Discord
Better for communities / dev groups.

📧 Email (for people who care about speed)

Superhuman
Expensive, but extremely fast if you process a lot of email.

🧰 Utilities that quietly carry

Bartender
Hides clutter in the menu bar.

Rectangle
Snap windows properly (what macOS should have natively).

💸 One tool I didn’t expect to need (but now use daily)

If you use AI APIs (OpenAI, Anthropic, etc.), cost visibility is terrible.

I started using this:

👉 https://tokenbar.site

It sits in your Mac menu bar and shows:

live token usage
cost per session
daily spend

It’s useful if you:

run agents
test prompts a lot
leave things running in the background

Caught a few runaway processes already.

🧠 Final thought

The biggest shift I’ve noticed:

Tools are becoming AI-native
Menu bar apps are underrated
Simpler tools win over “all-in-one” platforms

Curious what others are using—anything I’m missing?

Why I Track My AI Usage Like I Track My Finances

Michel Johnson — Tue, 03 Mar 2026 05:02:30 +0000

I've been using AI coding tools for over a year now. Claude, GPT, Codex, Cursor, Copilot — you name it, I've probably burned through a rate limit on it.

And for most of that time, I had no idea where my usage was actually going.

The Wake-Up Call

One Tuesday afternoon, mid-sprint, Claude just... stopped. Rate limited. No warning, no countdown, just a wall. I had maybe 3 hours of focused work left and my primary tool was gone.

So I switched to Cursor. Hit the limit there too within an hour — turns out I'd been burning through it all morning without realizing.

I sat there staring at my screen thinking: I pay for 5 different AI subscriptions and I have zero visibility into any of them. If this were my bank account, I'd be overdrafting every week.

The Financial Analogy

Think about it. Most of us track our finances somewhat carefully:

We check our bank balance before big purchases
We know roughly when bills are due
We set budgets and try to stick to them
We get alerts when balances are low

But with AI tools? We just... use them until they break. Then we panic, switch to another tool, and repeat the cycle.

The thing is, AI usage IS a financial decision. Between ChatGPT Plus, Claude Pro, Cursor, Copilot, and API credits, I'm spending $100-200/month on AI tools. That's a real line item. And the "currency" isn't just dollars — it's rate limits, session windows, weekly caps, and credit balances that all reset on different schedules.

What Changed

I started treating my AI usage like a budget:

Know your balances — At any given moment, I know roughly where I stand on each provider. How much of my weekly Claude limit is left. Whether my Codex sessions are trending hot.
Understand reset windows — Just like knowing when your paycheck hits, knowing when your limits reset changes how you spend. If Claude resets in 2 hours, maybe I save that complex refactor instead of burning through Cursor.
Pace yourself — This is the big one. If you know you're 60% through your weekly limit on Tuesday, you can ease up. If you're only at 20% by Thursday, you can go hard.
Have a fallback plan — When one provider is tapped, knowing which others have headroom means you're never fully blocked.

The Tool That Made It Click

I found TokenBar a while back — it's a macOS menu bar app that tracks all of this in one place. Usage limits, credit balances, reset timers, pace intelligence across 20+ providers. One glance and I know exactly where I stand.

What sold me was the pace tracking. It literally tells you if you're on-pace, in deficit, or have reserve relative to your reset window. It's like a budget tracker that says "hey, you're spending too fast this week" but for AI tokens.

$4.99 one-time (not a subscription, which is refreshing when everything else is monthly). Local-first, no telemetry.

The Mindset Shift

Once you start tracking, you realize a few things:

You waste more than you think. Half my Claude usage was re-explaining context I could have kept in a system prompt.
Timing matters. Some tasks are "expensive" (long code generation) and some are "cheap" (quick questions). Budgeting means doing expensive tasks when you have headroom.
Multi-provider is a strategy, not a fallback. When you can see all your balances, you can intentionally route work to the provider with the most capacity.

TL;DR

If you're spending real money on AI tools and hitting rate limits regularly, you owe it to yourself to track your usage the same way you track your spending. Know your balances, understand your resets, pace yourself, and have fallbacks ready.

Your future self (the one who doesn't get rate-limited during a critical deploy) will thank you.

What about you? How do you manage usage across multiple AI providers? I'm curious if others have found good workflows for this.

Anthropic Removed Claude's Usage Bars. Here's the Fix.

Michel Johnson — Tue, 03 Mar 2026 00:32:24 +0000

Anthropic removed Claude's usage progress bars. Here's what to do about it.

If you use Claude Pro or Claude Code, you've probably noticed: the session and weekly usage indicators are gone. No warning, no explanation. One day you could see your limits, the next day you couldn't.

The Problem

Without usage visibility, you're flying blind:

No idea how close you are to hitting the session limit
No idea when your weekly limit resets
No way to pace yourself across a workday
Surprise rate limit hits mid-conversation

This is especially painful if you also use Codex, Cursor, Gemini, or Copilot — each with their own invisible limits.

The Fix

TokenBar is a macOS menu bar app that puts your AI usage back where you can see it.

One icon. All your providers. Real-time limits, reset countdowns, and pace intelligence.

$4.99 one-time. No subscription. Local-first.

What It Tracks

Claude session + weekly limits
Codex usage percentage
Cursor monthly requests
Gemini RPM limits
Copilot allocations
20+ providers total (including Ollama for local models)

Why This Matters

The trend is clear: AI providers are making usage LESS visible, not more. OpenAI's Codex shows a vague percentage. Cursor buries usage in settings. Now Anthropic removed the bars entirely.

If you depend on these tools for work, you need external visibility. Don't let providers control what you can see about your own usage.

TokenBar — $4.99 one-time. macOS 14+.

How I Manage 6 AI Tools Without Losing My Mind

Michel Johnson — Mon, 02 Mar 2026 23:43:10 +0000

If you're using multiple AI tools for development in 2026, you already know the pain: different rate limits, different reset windows, different pricing models, and zero unified visibility.

Here's my actual daily setup and how I manage it without losing my mind.

The Tools

Tool	What I Use It For	Limit Type
Codex	Autonomous multi-file tasks	Session + weekly
Claude Code	Architecture, reasoning, code review	Session + weekly
Cursor	In-editor autocomplete + chat	Monthly requests
ChatGPT	Quick questions, brainstorming	Rate-limited
Gemini	Large context research	RPM limits
Copilot	Background completions	Included in plan

The Problem

Each tool has its own dashboard (if it even has one). Claude recently removed their usage progress bars. Cursor's usage page is buried in settings. Codex shows a vague percentage. None of them talk to each other.

When you're deep in a coding session and hit a rate limit on your primary tool, you lose 5-10 minutes figuring out which backup tool has capacity and when the primary resets.

The Solution

I found TokenBar — a macOS menu bar app that monitors all of these in one place:

Session/weekly/credit meters with reset countdowns
Pace intelligence — tells you if your usage rate will get you through the reset window
Status monitoring — incident badges when a provider is having issues
Threshold alerts — notifications before you hit limits
20+ providers supported (including Ollama for local models)

$4.99 one-time. No subscription. Local-first, no telemetry.

My Workflow Now

Morning: Glance at TokenBar. All green? Go hard on Codex for the big tasks.

Midday: Codex running low? Shift to Claude Code for reasoning work. Use ChatGPT for quick questions to preserve Claude capacity.

Afternoon: Check pace indicators. If I'm on track, keep going. If I'm in deficit, switch to Cursor for in-editor work (different quota pool).

Result: Zero surprise rate limits in 3 weeks. I always know exactly where I stand.

The Meta-Lesson

The bottleneck in AI-assisted development isn't the AI. It's the human overhead of managing multiple AI tools. Any system that reduces that overhead — whether it's an app like TokenBar or just a spreadsheet — pays for itself in recovered focus time.

What's your multi-AI management setup? Still winging it or have you built a system?

I Hit Rate Limits on 3 AI Tools in One Afternoon. Heres How I Fixed My Workflow.

Michel Johnson — Mon, 02 Mar 2026 17:41:18 +0000

Every morning I open my laptop and reach for 4 different AI tools before I even start coding. Claude for architecture thinking. Cursor for the actual coding. ChatGPT for quick questions. Copilot running in the background.

Last week I hit rate limits on THREE of them in a single afternoon. That's when I realized I needed to rethink my entire AI workflow.

The Multi-AI Problem

Here's the reality of being a developer in 2026: no single AI tool is good enough for everything. Claude reasons better but has strict limits. ChatGPT is fast but shallow. Cursor is great in-editor but limited outside it. Copilot autocompletes well but can't hold a conversation.

So you use all of them. And then you hit the management problem:

When does each limit reset?
How much capacity do I have left?
Should I use Claude for this task or save it for later?
Which tool is my fallback when the primary goes down?

What Actually Helped

I found TokenBar — a macOS menu bar app that monitors usage across 20+ AI providers. One glance tells me what I need to know: green means go, yellow means slow down, red means switch tools.

The feature that changed my workflow: pace intelligence. It doesn't just show how much I've used. It calculates whether my current rate will get me through the reset window. That's the actual useful information.

$4.99 one-time. No subscription. No telemetry. Local-first.

My Daily Routine Now

Morning check (2 seconds): Glance at TokenBar. Everything green? Go hard.

Midday adjustment: If Claude is running hot, shift lighter tasks to ChatGPT. Save Claude for the complex stuff.

Afternoon: By this point I know exactly where I stand. No surprises. No lost flow states.

Result: Zero surprise rate limits in 3 weeks.

The Bigger Lesson

The bottleneck in AI-assisted development isn't the AI. It's the human overhead of managing multiple AI tools. Any tool that reduces that overhead pays for itself immediately.

What does your multi-AI management look like? Still winging it or have you built a system?

I Built a Focus App Because Every Distraction Blocker Failed Me

Michel Johnson — Mon, 02 Mar 2026 17:40:06 +0000

I've tried every focus app on macOS. Freedom, Cold Turkey, SelfControl, Focus, One Sec — all of them. None of them stuck. Here's why, and what I built instead.

The Problem With Existing Focus Apps

Every distraction blocker I tried had the same fundamental flaw: they treat focus as a binary state. You're either "blocked" or you're not. But real focus doesn't work that way.

Here's what actually happens during deep work:

You start working
Your brain gets an urge to check something (Twitter, email, Slack)
If the app blocks it, you feel frustrated and find a workaround
If the app lets you through with a "are you sure?" prompt, you click yes every time
Either way, the focus is broken

The issue isn't blocking. It's awareness.

What I Built

Monk Mode takes a different approach. Instead of just blocking, it tracks every time your brain tries to distract you during a focus session. Every URL you try to visit, every app you try to open — it logs the impulse.

At the end of a session, you get a distraction report:

14 times you tried to open Twitter
7 times you tried to check email
3 times you tried to open YouTube

Seeing the raw numbers is more powerful than any block. It creates self-awareness about your distraction patterns that no blocker can provide.

The Stack

SwiftUI for the UI
macOS Accessibility APIs for app monitoring
Network Extension for URL tracking
Local-first — no data leaves your machine

Key Design Decisions

Blocking + awareness, not just blocking. The app does block distractions, but the awareness layer is what actually changes behavior over time.

Session-based, not always-on. You start a session when you want to focus. No persistent background monitoring.

No subscription. $15 lifetime, or $10 with code DEV. I hate subscription fatigue for utility apps.

Results

After using it myself for 3 months:

Average focus session went from 25 min to 52 min
Distraction attempts dropped from ~40/session to ~12/session
The awareness effect compounds — you start catching yourself BEFORE reaching for the distraction

What I Learned Building It

macOS security APIs are painful. Network Extensions require special entitlements and Apple review. Took 3 weeks just to get approval.
Users want data, not just restrictions. The distraction report is the #1 feature people mention in reviews.
Simplicity wins. Every feature I removed made the app better.

If you struggle with focus during coding sessions, check it out at mac.monk-mode.lifestyle. Code DEV gets you $5 off.