CodexZero CLI
A Codex wrapper that folds repeated command output before the model reads it and, by the author's tests, saves up to 20% of tokens
- 80 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Installation runs a script downloaded over the network and creates a separate copy of the Codex app with a patched settings tab
- The output of all commands passes through the processor and provider API keys are stored in the app
Install
- Works in
Manual install
irm https://raw.githubusercontent.com/Retro2512/CodexZero/main/scripts/bootstrap.ps1 | iexWindows: downloads the installer, verifies it and installs the app with no admin rights.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Installation runs a script downloaded over the network and creates a separate copy of the Codex app with a patched settings tab
- The output of all commands passes through the processor and provider API keys are stored in the app
This is third-party code. Review the repository files before installing.
What it does
CodexZero is a layer over Codex that removes repeats from command output before the model reads it. Build logs, test runs and file dumps often print the same line many times, and tokens are charged for every copy. CodexZero saves the full output to disk, collapses identical lines into a count and hands Codex a short version: errors and anything new stay. The model and answers stay the same. Three console commands work: `codex-zero run` starts Codex in the lean mode, `codex-zero savings` shows your savings and `codex-zero stock` starts regular Codex. Modes are standard, focused and safe. The CodexZero app adds custom models in the shared picker, a live chat cost and a cache indicator.
Who it is for. Codex users who run long builds and tests and hit token limits, and those who want to see the cost of every chat.
Good fit when
- The agent runs builds and tests and one warning repeats dozens of times in the log
- You want `codex-zero savings` to show how many tokens your own work saved
- You want to add a Claude, Z.ai or local model key to the Codex model list
- During work you want live chat cost and a cache warmth gauge
- The task needs many independent tool calls and the focused mode fits
Not a fit when
- You need plain Codex with no output processing: `codex-zero stock` exists for that
- You work on Linux: the lighter app is built only for Windows and on macOS you install the app or the command
- Command output is short with no repeats: there is nothing to save
- You want to connect a Claude Code subscription: the project connects provider APIs, not login sessions
Example request
Run this project through codex-zero run, run the tests and show with codex-zero savings how many tokens were savedLimitations
The savings figures were measured by the author on DeepSWE and Terminal-Bench against regular Codex: the Standard mode gave 17.1% fewer tokens with every task passed. On your work the result depends on how many repeats the output has. Installation uses the `bootstrap.ps1` script through `irm | iex` on Windows and `bootstrap.sh` through `curl | sh` on macOS, which downloads and verifies the installer. The lighter app with slides, docs and plugins removed is Windows only for now. Claude and Z.ai requests go under your API key, and keys are DPAPI-encrypted on Windows.
How to disable. On Windows uninstall the app through Settings, Apps and on macOS drag it to the Trash. The rollback guide is in `docs/rollback.md`. The installed Codex, login and settings are not modified, and the regular shortcut starts the regular app.
FAQ
How do I install CodexZero?
On Windows run `irm https://raw.githubusercontent.com/Retro2512/CodexZero/main/scripts/bootstrap.ps1 | iex` and on macOS `curl -fsSL https://raw.githubusercontent.com/Retro2512/CodexZero/main/scripts/bootstrap.sh | sh`. For the console command on macOS replace `| sh` with `| CODEX_ZERO_INSTALL=cli sh`.
Does the model lose information from the output?
The full output is saved to disk and Codex gets a short version with errors and new lines. If the original is shorter than the compressed one, the model reads the original.
How do the modes differ?
`standard` is the default with short instructions and normal Codex tools, `focused` batches independent tool calls and `safe` uses regular tools and instructions.
How do I go back to regular Codex?
Run `codex-zero stock` or open the regular Codex shortcut: the CodexZero app runs alongside and does not change your Codex, login or settings.
Can I connect other models?
Yes, the app settings have a Custom models section: Claude, Z.ai Coding or your own model with a model ID and API key. OpenAI Responses, Chat Completions and Anthropic Messages are supported.
Discussion
No questions or reviews yet.
Related
CLI toolsAn unofficial Codex CLI build with rewind: /rewind returns the conversation and project files to a chosen prompt, /redo undoes it.
Codex Settings skills
Codex CLI Skills & Profiles
A Codex CLI plugin with twelve skills: deep research, pull request review, image generation and session handoff
A Codex skill that runs a loop of measurable experiments over a repository toward a numeric target: change, verify, revert failures, repeat
A Codex plugin that turns local task logs into a timeline with token counts, tool timing and estimated cost for every step.
Codex Trace reads local Codex CLI logs and shows turns, tool calls and token usage in a desktop window or in the browser.
A layer on top of OpenAI Codex that adds project memory, planning and completion checks before a task counts as done
README in short
The README explains that repeats are compressed before the model reads the output and shows a benchmark table against regular Codex and other tools. It then covers installation for Windows and macOS, three commands, app extras, modes and compatibility.