Code
A CLI that understands your browser. Sign in with ChatGPT!
Start new thread

Code - A CLI that understands your browser. Sign in with ChatGPT!

by•
Fast, effective, mind-blowing, coding CLI. Browser integration, multi-agents, theming, unified diffs, and reasoning control. Orchestrate agents from OpenAI, Claude, Gemini or any provider. npm install -g -every/code code // or coder if you have vscode

Add a comment

Replies

Best
Maker
📌
Apologies for the terrible video. That's what a midnight editing session gets you! I love coding with GPT-5, but codex... needs a bit of work. After submitting a few PRs, I decided it would take too long for it to become the tool I wanted, so I forked and implemented some major upgrades. I have been working on various MCPs with limited versions of these concepts, but all were impossible to implement fully without having control over the CLI agent. Allowing GPT-5 to act in a more agentic environment really shows it's power. I mostly built this for myself, but thought I'd throw this out there to see if it hits a need anyone else has. Fully open source. Feel free to fork, or submit PRs. I promise to be more responsive than OpenAI :)

 Great work James keep it up (Y)

Maker

 Thanks!

I love the concept of unified diffs and theming it definitely makes debugging a lot cleaner. Is there a way to save and share custom themes or diff setups for the team?

Maker

 

Thank you for the kind words! You can absolutely create your own themes. Code’s theme system lets you start from a base theme and override colours or build an entirely custom palette via your `~/.codex/config.toml`. To share with teammates, you can commit your theme section into your repo and have others copy it into their `.codex` directory. Diff display settings live in the config too; there’s no built‑in cloud sync yet, but we’re exploring easier ways to share presets.

The reasoning control is powerful. Can we set custom prompts for each agent?

Maker

  Appreciate it! You can adjust reasoning effort globally via the `/reasoning` command or configure different models and settings via profiles in `config.toml`. At the moment there’s no per‑agent custom prompt hook – multi‑agent commands like `/plan` and `/code` construct prompts automatically – but you can influence behaviour by putting guidance in an `` file. Fine‑grained per‑agent prompts are something we’re discussing internally.

Does the CLI allow real-time collaboration between agents or is it designed more for executing tasks one after the other?

Maker

  Right now Code is focused on single-developer workflows. There isn't a real-time collaboration mode between human users; the multi-agent commands run GPT-5, Claude and Gemini together on your machine but don't sync sessions across developers. For now you share diffs via version control. Real-time collaboration is on our radar.

Can the CLI be fully integrated into CI/CD pipelines or is it primarily meant for local development?

Maker

  Yes! Code includes a non-interactive mode for CI/CD. You can pass prompts directly with flags such as --no-approval or --quiet to run it headless. The docs include an example GitHub Action that updates changelogs or runs tests in a pipeline.

This CLI looks super versatile with its multi agent orchestration. Can you link reasoning across different providers like OpenAI → Claude → Gemini all in one workflow?

Maker

  Definitely. Code is built for cross‑provider orchestration. You can install Claude and Gemini adapters alongside Code and then use multi‑agent commands to coordinate them. For example, `/plan` asks GPT‑5, Claude and Gemini to build a shared plan, `/solve` runs a race between them and picks the fastest valid answer, and `/code` applies their consensus before making changes. We’re still working on deeper context passing, but multi‑provider workflows are supported today.

This is absolutely brilliant. Is it possible to set up which agent automatically manages specific file types or tasks?

Maker

  Thanks! Right now you choose the model with a flag or multi-agent commands. There's no way to auto-assign by file type yet, but you can run tasks separately. Assigning models to file types is a great suggestion we're exploring.