Which Terminal AI Agent Ought to You Use?

Coding assistants have moved past autocomplete into full brokers that may learn initiatives, run instructions, edit information, and iterate towards outcomes. Instruments like Claude Code and Codex each function on this area, however take totally different approaches. Claude Code facilities on a unified agent loop throughout environments, whereas Codex spreads capabilities throughout CLI, IDE extensions, cloud workflows, and delegated duties.

This isn’t about mannequin efficiency. It’s about workflow: management, intuitiveness, and the way simply you possibly can keep targeted whereas working inside an actual repository. On this article, we evaluate how every software suits into the act of getting work executed.

Getting began with Claude Code and Codex CLI

Earlier than shifting onto the true workflows, First let’s set up each the instruments in our system. Please be certain your system has node already put in. 

Codex CLI 

Set up the Codex CLI with npm. Open your terminal and run 

npm i -g @openai/codex

Run Codex in a terminal. It will probably examine your repository, edit information, and run instructions. 

Codex 

Check in with an OpenAI account or API key

Claude Code 

Set up the Claude Code with npm. Open your terminal and run 

npm set up -g @anthropic-ai/claude-code 

Run in terminal by altering the listing to specific venture 

claude 

Check in with an Anthropic Account

Now all set, let’s transfer to workflows.

The primary 10 minutes really feel totally different

Claude Code seems like an assisted accomplice. It needs to get a deal with on the repo, recommend a plan, then proceed with the duty with mode permission and checkpoints to maintain it secure. Codex seems like a configurable runtime. It’s nonetheless conversational, however the focus is extra on configuration, insurance policies, worktrees, overview, and cloud delegation. 

In case you are opening a repo for the primary time, the hands-on distinction exhibits up instantly. 

With Claude Code, a pure first transfer is:

Clarify the auth circulate, record the dangerous information, and inform me the place login may very well be failing.

With Codex, the equal seems like: 

Clarify the auth circulate, record the dangerous information, and inform me the place login may very well be failing 

The identical immediate, however the expertise could be very totally different. Claude usually encourages you to plan and execute. With Codex it feels prefer it asks you to set the parameters of freedom, sandboxing and approvals earlier than leaping in. 

That distinction issues. Should you like being guided to productiveness, you’ll like Claude Code extra. Should you prefer to design a system, Codex is extra rewarding. 

The Translation Layer: How the ideas map?

A lot of the confusion of Claude Code vs Codex is because of totally different terminology. 

That is the mannequin to remember whenever you learn the remainder of this text. 

Repo directions: CLAUDE.md vs AGENTS.md

It is a essential a part of the article as a result of it impacts how the agent feels after the primary day. 

Claude Code masses CLAUDE.md in the beginning of every session and makes use of it as context for the venture, your Workflow, and even your organization. Anthropic’s documentation is evident that you must use CLAUDE.md to seize the principles you don’t need to repeat, and use auto reminiscence for Claude’s studying.  

The Codex answer makes use of AGENTS.md, however in a extra refined means. You can have a world ~/.codex/AGENTS.md, then AGENTS.md per repo, then sub AGENTS.override.md, all as a part of the config.toml construction.  

Right here’s the way it may work. 

Right here’s a helpful CLAUDE.md for a Node repo: 

A helpful AGENTS.md for a similar repo may seem like this: 

The hands-on lesson is easy. Don’t wait till the agent disappoints you 5 instances. Write the instruction file early. Each instruments get a lot better as soon as your requirements reside within the repo as a substitute of in your head. 

Reminiscence: What will get remembered and the way helpful it truly is?

The context window for Claude Code is wiped initially of every session, however you possibly can load your CLAUDE.md and auto reminiscence. Based on Anthropic, auto reminiscence is notes that Claude writes based mostly in your corrections and preferences, resembling construct instructions, debugging hints and issues it has seen whereas modifying in that tree.  

Codex Recollections are related however they’re barely extra specific. Recollections are disabled by default, are saved domestically (in ~/.codex), and are for mounted preferences, widespread routines, project-specific conventions, and customary gotchas. The OpenAI docs additionally advise to not retailer reminiscences of guidelines as the one place for guidelines that should at all times be adopted. These nonetheless have to go in AGENTS.md or in paperwork within the repo.  

This leads to an awesome workflow. 

In case you are utilizing Claude Code, you possibly can have the agent study the tempo of the repo, then use CLAUDE.md for issues it is advisable to maintain secure. 

In case you are utilizing Codex, don’t put the contract in Recollections. Put the contract in AGENTS.md. Put your platform guidelines in config.toml. Let reminiscences fill within the gaps. 

This makes Codex really feel extra mechanical. Claude is extra like a wise teammate. 

Permissions and planning: That is the place the character cut up turns into apparent

Claude Code has very descriptive names for permission modes. The obtainable modes are at the moment default, acceptEdits, plan, auto, dontAsk, and bypassPermissions. plan is especially fascinating because it permits Claude to plan and suggest adjustments with out touching your supply, and auto is a analysis preview that makes use of an additional classifier to filter actions.  

Codex describes this by way of sandbox and approval coverage. OpenAI’s documentation calls sandbox mode the technical sandbox and approval coverage the rule for when to ask permission. Native Codex by default makes use of no networking and sandboxing underneath the OS, which is generally configured by way of ~/.codex/config.toml and, optionally, project-specific .codex/config.toml.  

Right here is the hands-on model. 

If you would like Claude Code to examine a repo and produce a proposal earlier than touching something: 

claude --permission-mode plan 

If you would like Claude Code to maneuver quicker on secure file edits: 

claude --permission-mode acceptEdits 

If you would like Codex configured for a tighter read-only go first, the OpenAI docs present patterns like this: 

Open the .codex/config.toml file and add the next traces:

[profiles.readonly_quiet] 
approval_policy = "by no means" 
sandbox_mode = "read-only"

Then you should utilize that form of profile for a first-pass audit and solely loosen up it when you find yourself prepared. 

This distinction issues so much in actual groups. Claude exposes the security mannequin as an interplay sample. Codex exposes it as a system configuration sample. 

Let’s say your checkout check is failing and also you need the agent to analyze, repair, confirm, and clarify the change. 

An excellent Claude Code workflow seems to be like this: 

Discover why the checkout is failing. Begin in plan mode, determine the smallest secure repair, implement it, run the related assessments, and summarize the change in plain English. 

An excellent Codex workflow seems to be like this: 

Examine the checkout failure, maintain scope minimal, clarify root trigger first, then patch solely the information required, run the smallest related check set, and present me the diff I ought to overview. 

Discover the distinction. With Claude Code, you naturally lean into circulate. With Codex, you naturally lean into specific scope and overview language. 

Each instruments can do the loop, however they encourage barely totally different types of prompting. 

Undo, restoration, and reviewing adjustments

Claude Code’s undo/rewind is a robust characteristic. Anthropic claims that each user-prompted change makes a checkpoint, the checkpoints are persistent, and /rewind can restore code, dialog, or each. So you possibly can “experiment” extra with out worrying about errors.  

A “actual” use case seems to be like this: 

/rewind 

You then select whether or not to simply rewind the code, simply the chat, each, or begin summarising from a specific level and proceed. 

And Codex addresses security in one other means. The overview pane shows the adjustments within the repo, means that you can add inline feedback and to stage, maintain or revert traces. The app additionally makes use of worktrees so many issues can occur when you work in your checkout. 

So the sensible cut up is that this: 

Claude says, “Strive the dangerous factor. You possibly can rewind.” 

Codex says, “Let the work occur in isolation. Then examine it fastidiously.” 

Each are good. They simply change how daring you are feeling whereas iterating. 

Abilities, hooks, and reusable workflows

That is the part the place superior customers begin constructing actual leverage. 

Claude Code expertise use SKILL.md, and Anthropic claims Claude can robotically invoke expertise as wanted, or you possibly can explicitly use slash instructions (e.g. /review-pr or /deploy-staging). Claude additionally has hooks for operating shell instructions earlier than or after Claude Code actions, resembling formatting, linting or customized validation.  

OpenAI’s docs for Codex give attention to progressive disclosure. Codex masses ability metadata and solely masses the complete SKILL.md when it makes use of the ability. Codex additionally makes use of a built-in $skill-creator, and has hooks as an experimental extensibility framework (characteristic flag is in place). 

Here’s a concrete hands-on sample you should utilize in both software. 

Create a reusable code-review ability that claims: 

--- 
identify: backend-review 
description: Evaluate backend adjustments for auth bugs, migration threat, logging gaps, and check protection regressions. 
---

When invoked: 

1. Examine modified information first 
2. Prioritize auth, knowledge integrity, and silent failure modes 
3. Recommend the smallest fixes 
4. Finish with a brief threat abstract 

In Claude Code, that turns into one thing you possibly can naturally name from the dialog. In Codex, that turns into a cleaner reusable unit in a extra explicitly managed system. 

Which one must you select?

Primarily based of the comparability and the options the 2 supply, right here’s a comparability desk to summarise all of it:

Conclusion

Claude Code wins on simplicity and “circulate.” The /rewind characteristic is a top-tier security web. The auto-memory system makes it really feel sensible over time. Select Claude Code if you would like aPair Programmer that simply works. It’s glorious for fast prototyping and refactoring. 

Codex wins on precision and configurability. The worktree mannequin is ideal for complicated automation. The policy-based permissions swimsuit enterprise safety wants. Select Codex if you wish to construct a customized platform. It’s a strong alternative for systematized improvement. 

These instruments usually are not simply opponents. They characterize totally different futures for AI coding. One is a guided agent. The opposite is a programmable runtime. They’re catered to totally different customers and each help in bettering your workflows.

Continuously Requested Questions

Harsh Mishra is an AI/ML Engineer who spends extra time speaking to Massive Language Fashions than precise people. Keen about GenAI, NLP, and making machines smarter (in order that they don’t substitute him simply but). When not optimizing fashions, he’s most likely optimizing his espresso consumption. 🚀☕