Claude Code vs Cursor vs Codex

Claude Code vs Cursor vs Codex is a question about where you want to sit while an agent edits your repo. Claude Code is Anthropic's terminal agent.

Cursor is an AI-first editor with Agent/Composer. Codex is OpenAI's coding agent, the one I mean in CLI and workspace form, not the 2021 autocomplete model.

I already split these into pairs: Claude Code vs Cursor and Claude Code vs Codex. This is the three-way for people who refuse to open two tabs.

What is the real split in Claude Code vs Cursor vs Codex?

Two of them want a tty. One of them wants a window.

That split decides review, not intelligence. I trust Cursor more when I want to click through a diff with my eyes.

I trust Claude Code more when I already live in tmux and I will read git diff anyway. I trust Codex when the rest of my work is already in OpenAI's world and I want that CLI, not a second vendor habit.

All three can use tools. All three can run tests.

All three can waste an afternoon on the wrong file.

Claude Code

Surface: terminal

Feel: plan, tool, patch, ask

MCP: first-class; claude mcp add --scope user …

Best for: people who already review with git

Cursor

Surface: VS Code fork

Feel: chat + inline + Agent, visual diffs

MCP: mcp.json, one-click from some apps

Best for: people who want the file tree in the same glass

Codex

Surface: CLI / workspace agent (see Codex CLI)

Feel: OpenAI-shaped agent loop

MCP: supported in current builds; setup is CLI-first

Best for: teams standardized on OpenAI

Surface is the product more than model brand. You can put more than one model behind a surface.

You cannot put a terminal habit into a person who will not read diffs.

How do I actually work in each one?

Claude Code: I cd to the repo, I state the constraint, I let it run, I read the diff, I / my way through the session if I need to. How to use Claude Code is the on-ramp.

Cursor: I open the file I think is true, I @-mention it, I let Agent propose, I accept hunks. I start a new chat when the thread is junk.

Codex: I treat it like Claude Code with a different login and a different set of sharp edges. I do not assume the flags I learned in claude exist.

None of these steps include "look at the running UI." That is the shared hole.

What do they cost, roughly?

Cursor Pro is typically around $20/month with a fast-request pool you can burn in agent-heavy weeks.

Claude Code rides your Anthropic / Claude plan and usage. Heavy repo work is not a flat feeling even when the sticker looks simple.

Codex rides your OpenAI plan and whatever workspace limits exist that month.

I will not invent a tokens-per-dollar table. Prices move.

The honest cost is "I already pay vendor X, so I start there."

A one-time Mac capture license is a different line from a third seat. Pricing.

Can any of them see the element I mean?

Not natively.

They see the repo, the terminal, tool output, and whatever image you paste. A PNG is not AXButton "Publish".

PinVari fills that for all three. ⌥⌘A to mark, ⌥⌘V voice-only, ⌥⌘P pause on scroll.

On-device STT and OCR. Local MCP at ~/.pinvari/mcp/pinvari-mcp talking to 127.0.0.1:3402.

claude mcp add --scope user pinvari -- "$HOME/.pinvari/mcp/pinvari-mcp"

Cursor and VS Code: PinVari → Connect. Codex: same connector, CLI config.

Never bare claude mcp add pinvari.

pinvari_next_instruction returns path, words, crop, confidence, circled-versus-dwelled. pinvari_request_capture lights the notch island mid-task.

pinvari_mark_done closes.

On Electron, the tree may be lazy. The capture side sets AXManualAccessibility and retries.

On a bare AXGroup, it descends to a labeled child. Below ~0.8 confidence it should ask.

Named context transfers across vendors. That is why I will not pick a winner for you.

I will pick a capture format.

How should a Claude Code regular choose?

Stay in Claude Code if you already have a ritual and you review with git. Add MCP capture.

Read Claude Code vs Cursor only if you miss inline diffs.

Try Cursor for a week if you keep losing the file in the terminal scrollback. You can still call Claude models inside Cursor.

The surface changed, not your brain.

Try Codex if your org already bought that stack and Claude is the exception. Do not run Codex and Claude Code on the same dirty working tree at the same time.

That is how you get two agents politely fighting.

I have done that. It is funny until it is a merge.

Decision cards

Want visual diffs, stay in a window: Cursor

Want tty, Anthropic-native tools: Claude Code

Want tty, OpenAI-native tools: Codex

Want the right button edited: capture + any of the three

What do I keep in the repo so all three behave?

A short rule: stack, test command, "do not edit files I did not name."

A way to run the app. If the agent cannot start the preview, it cannot verify a visual fix even with a perfect capture.

One MCP capture server, not two screenshot toys.

Git, committed before you let any of them off the leash.

If you work across monitors, the mark must remember the display. If you review long pages, pause/resume and scroll-aware frames matter more than which vendor logo is in the title bar.

Focused-window text up to 40,000 characters exists as an option. I do not lead with it.

I lead with the mark.

The public tool list at pinvari.com/mcp is what I want every agent-facing server to publish. Claude Code, Cursor, and Codex all consume the same shape.

I will not crown a winner. I will say: pick the surface you will actually review on, then stop describing buttons in prose.

How I migrate a habit from one of these to another

I do not export chat history. I export the constraints.

I copy the three-line rule file, the MCP config, and one example capture that worked. I throw away the thread where the agent was wrong for an hour.

That thread is not an asset.

In Cursor I recreate rules under .cursor/rules. In Claude Code I keep a CLAUDE.md or the project memory I already trust.

In Codex I use whatever instruction file that CLI actually reads this month, and I verify with a dry prompt before I trust it.

I run the same visual bug once as a bake-off. Same sentence, same mark, same repo state.

I score right file, right control, number of follow-ups. I do not score vibes.

If Codex wins on a Tuesday and Claude Code wins on a Wednesday, I stay with the one whose diffs I already review faster. Oscillating vendors is a hobby.

Shipping is a habit.

Permissions stay on the Mac either way. Accessibility and Screen Recording belong to the capture app, not to three agents.

I do not grant screen recording to every CLI because a README said "computer use."

When a vendor adds a pixel-clicking demo, I treat it as a demo. Named AX plus a crop is still the thing I will send on a Friday afternoon anyway.

FAQ

Is Claude Code better than Cursor in 2026?

For terminal-native developers, often yes. For people who review in a visual diff UI, Cursor.

Model quality is close enough that the surface decides the day.

Where does Codex sit in Claude Code vs Cursor vs Codex?

With Claude Code on surface, with OpenAI on billing and defaults. Compare it to Claude Code first.

Compare that winner to Cursor second.

Can I use the same MCP server in all three?

Yes if it is a real MCP server on localhost. Connectors differ.

The payload should not.

Which one wastes fewer tokens on screenshots?

None of them are cheap if you paste a desktop. A named crop is cheaper in every client.

Token waste is an input problem.

Should I run two agents at once?

No, not in one dirty tree. Serial is fine.

Parallel is a merge tax.

Do I need this comparison if I only file tickets?

Only the capture half. The tracker does not care which agent you would have used.

The engineer or the later agent will.