My AI coding workflow on a Mac in 2026, step by step: 8 tools from prompt to review
DEV Community

My AI coding workflow on a Mac in 2026, step by step: 8 tools from prompt to review

You open a terminal, type a prompt, and twenty minutes later an agent hands you a diff. Everything around that moment (where you type, where you read, how you check the agent didn't wreck anything) decides whether the day feels fast or chaotic.

Here's the AI coding workflow I'd set up on a Mac in 2026, step by step, with the one tool I'd use for each step.

The loop Prompt, run, review, fix, repeat. Each tool below owns exactly one part of that loop. If a tool doesn't make one step faster, it doesn't make the list.

Step 1: The terminal where agents live → Ghostty

Ghostty is a fast, native terminal emulator, free and open source. Coding agents spend most of their life printing to a terminal, so this is the window you'll stare at most. Splits and tabs are built in, which is handy for one agent per pane.

Step 2: The agent that does the long jobs → Claude Code

Claude Code reads your codebase, edits files and runs commands. It runs in the terminal, IDE extensions, a desktop app and the web.

Install on macOS with:

curl -fsSL https://claude.ai/install.sh | bash

Most surfaces need a Claude subscription or an Anthropic Console account.

Step 3: The second agent → Codex CLI

Codex CLI is OpenAI's coding agent that runs locally on your computer. It's open source.

Install with:

curl -fsSL https://chatgpt.com/codex/install.sh | sh

Having two agents means one can take a different task, or a second pass at the same one.

Step 4: Reading the diff → Cursor

Agents write fast; reviewing is where your time goes. Cursor is an AI code editor that's comfortable for reading changes across many files, and it has its own agent if you'd rather prompt from the editor. I'd keep it open next to the terminal purely as the review screen.

Step 5: Keeping both on screen → Rectangle

Rectangle moves and resizes windows with keyboard shortcuts or snap areas. Free, open source, macOS 14+. Terminal on the left half, editor on the right half, one shortcut each. That layout is the whole workflow.

Step 6: Writing long prompts without typing them → Wispr Flow

Good prompts are long. Wispr Flow is voice dictation that works across your apps, and dictation is free on Mac (paid plans add team and admin features). Talking through the task, the constraints and what "done" means is often quicker than typing it.

Step 7: Showing the agent what's broken → CleanShot X

When the bug is visual, a screenshot beats a paragraph. CleanShot X does screenshots, screen recording and annotation. The Basic license is $35 once (one year of updates), and it needs macOS 13 or newer. Annotate the broken spot, then drop the image into the agent.

Step 8: Knowing if you can afford the next run → TokenBar

The step people skip: checking how much limit you have left before starting a long task. Claude, Codex, Cursor, Gemini and Copilot all show usage in different places. TokenBar is a menu bar app that shows usage and reset times for those five in one place. It's built on Peter Steinberger's open source CodexBar (MIT) and credits Pixel Agents for its agent view. $5/month Basic or $15 lifetime Pro, macOS 14+. It doesn't give you extra quota; it just tells you where you stand.

The whole stack at a glance

Step Tool Price
Terminal Ghostty Free
Agent 1 Claude Code Claude plan or Console account
Agent 2 Codex CLI Open source CLI
Review Cursor Free and paid plans
Layout Rectangle Free
Prompting Wispr Flow Free dictation on Mac
Bug context CleanShot X $35 once
Limits TokenBar $5/mo or $15 lifetime

Which step of your loop is still the slow one: prompting, waiting, or reviewing?

Read on DEV Community ↗ ← Back to News

Comments

No comments yet. Start the discussion.