AI Development Environment: Cursor + Claude Code + Codex

A setup log documenting the installation and configuration of an AI-assisted development environment — Cursor IDE with the Claude Code and Codex coding-agent CLIs, wired to a GitHub workflow.

Tools Installed

Tool	Version	Notes
Cursor IDE	3.4.20	Free plan
Claude Code CLI	v2.1.142	Anthropic OAuth · Claude Max · Opus 4.7 (1M context)
OpenAI Codex CLI	v0.130.0	OpenAI account · gpt-5.4 medium
Node.js	v22.22.2	—
npm	10.9.7	—
Git	2.50.1 (Apple Git-155)	Bundled with macOS

Platform: macOS (Apple Silicon), zsh.

Steps Completed

1. Downloaded Cursor for macOS from cursor.com and installed via DMG.
2. Signed into Cursor with my GitHub account on first launch.
3. Opened the Extensions panel (⌘⇧X) and searched for "Claude Code" and "Codex" — neither appeared. Investigated, then realized both ship as npm-distributed CLIs, not VS Code marketplace extensions.
4. Opened Cursor's integrated terminal and installed Claude Code:

   npm install -g @anthropic-ai/claude-code

5. Launched Claude Code (claude), picked dark theme, completed browser OAuth at platform.claude.com → logged in as nabinkumar585@gmail.com. Came up with Opus 4.7 (1M context) on Claude Max.
6. Installed Codex:

   npm install -g @openai/codex

7. Launched Codex (codex). On first run it prompted to upgrade to the newly released GPT-5.4 model (with a tip that GPT-5.5 was also available). Accepted; landed on gpt-5.4 medium.
8. Created this public GitHub repository.
9. Committed and pushed this README via GitHub's web editor.

Issues Encountered and How I Solved Them

1. Claude Code and Codex aren't in Cursor's Extensions marketplace. The brief said to install them as Cursor extensions, but searching the Extensions panel (⌘⇧X) returned nothing relevant. Realized both Claude Code and Codex are distributed as npm-installable CLI tools, not VS Code/Cursor marketplace extensions. The IDE integration registers automatically when the CLI runs from inside Cursor's integrated terminal. Fix: ignored the Extensions panel; installed via npm install -g.

2. zsh: command not found: # errors from pasted commented commands. When I pasted a multi-line snippet that included # comment lines, zsh interpreted each # line as a literal command and printed command not found: #. The real npm install and claude commands underneath still ran — the errors were harmless noise. zsh in interactive mode doesn't treat # as a comment unless interactive_comments is set. Fix: ignored the errors once I understood the cause; lesson is to strip comments before pasting, or set the option in .zshrc.

3. Native Claude Code install was outside PATH. After authenticating, Claude Code warned:

Native installation exists but ~/.local/bin is not in your PATH.

Two claude binaries existed — the npm one (working) and a native install in ~/.local/bin (not on PATH). Fix: appended the directory to my shell config:

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && source ~/.zshrc

New terminal sessions now resolve the native binary too. (Worth noting: source only refreshed the current subshell — the permanent .zshrc change is what matters for future sessions.)

4. Picked the wrong Claude Code login method. Claude Code's login prompt offered three options: (1) Claude account with subscription, (2) Anthropic Console (API billing), (3) third-party platforms. I have Claude Max — option 1 would have used my existing subscription at no extra cost. I defaulted to option 2 (API), which bills my Anthropic Console credits. The cost is negligible (cents for this setup), but it's a real "read the options carefully" lesson. Would re-auth with option 1 next time.

What I'd Do Differently Next Time

• Skip the Extensions panel entirely for any AI coding tool — go straight to npm or the tool's official install script.
• Read login options carefully — pick the subscription-based flow over API billing when a subscription is available.
• Set interactive_comments in .zshrc so # lines in pasted commands are tolerated.

Notes on the Two Agents

Brief impressions from running both side-by-side during setup:

• Claude Code (Opus 4.7) — strong on multi-step instructions; visible reasoning; asks permission before running shell commands; clean in-terminal welcome with version info, current model, and shortcut hints.
• OpenAI Codex (GPT-5.4 medium) — similarly agentic; prompt shows model + working directory above the input; flagged its own model upgrade on first launch.

Both are CLI-first with IDE integration. The marginal cost of having both installed is zero — different agents, different strengths, worth keeping both around.