Skip to content
beanbocchiPublic

About

The pro(xy) version of your coding agent.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

10 Commits

Folders and files

Repository files navigation

agentpro

The pro(xy) version of your coding agent.

agentpro runs Claude Code and Codex CLI against a shared CLIProxyAPI server, so one API key reaches every model on it. The server owner hands out the keys; none is in this repository.

  • claudepro starts Claude Code.
  • codexpro starts Codex CLI.

Install

Linux and macOS, from bash, zsh or fish. Needs curl and python3:

curl -fsSL https://raw.githubusercontent.com/beanbocchi/agentpro/main/install.sh | bash

Windows, from PowerShell:

irm https://raw.githubusercontent.com/beanbocchi/agentpro/main/install.ps1 | iex

The installer asks for the key and checks it against the server before writing anything. It installs Claude Code if claude is missing; codexpro needs Codex CLI, which you install yourself. CLAUDEPRO_KEY, and CLAUDEPRO_URL for another server, skip the prompt.

agentpro updates itself. At each start, claudepro and codexpro check GitHub in the background, at most once an hour, and install the newest commit on main with the saved key; the next start prints agentpro: updated to <commit>. To update at once, run the installer again. An install made from a clone of this repository does not update itself, and AGENTPRO_NO_UPDATE=1 turns updates off for a run.

The commands land in ~/.local/bin on Linux and macOS, and in %USERPROFILE%\.claudepro\bin on Windows, which the installer adds to the user PATH.

Use

claudepro                        # same arguments as claude
claudepro -p "explain this repo"
codexpro                         # same arguments as codex
codexpro exec "explain this repo"

Neither asks for permissions: Claude Code runs with --dangerously-skip-permissions, Codex with approval_policy=never and sandbox_mode=danger-full-access.

In claudepro, /quota shows the plan limits of the server's accounts. /contribute logs in to a Claude account and adds it to the server (/contribute codex does the same for a ChatGPT account); it prints a login address, and after the login you paste back the address the browser landed on. /login and /logout are turned off there, since the session signs in with the server's key. The dim line under the prompt ends with the tokens the session and the day have used, and, for Claude and for Codex, how many accounts' worth of limit the live accounts have left (each counted by its tightest window, 5-hour or weekly). Ask for a picture and Claude calls the generate_image tool, which draws it with one of the server's image models and can edit or follow reference images. The tool is there only when the server lists an image model, and it needs Python (python3, or python on Windows).

The first time, Claude asks which image model to use, and that model becomes the default for every later session. /modelimg shows the image models and the default, and /modelimg <model> changes it; asking Claude in chat ("use gpt-image-2 from now on") works too. Naming a model in a request uses it for that picture only. In claudepro -p, with nobody to ask, a session without a default uses the newest image model.

In codexpro, Codex's own image tool draws pictures through the server, with no OPENAI_API_KEY. Codex picks the image model itself (gpt-image-2), so /modelimg applies to claudepro only.

Models

/model lists every chat model on the server, read again at each start, so a model the server adds shows up the next time. Image, audio and video generation models, embeddings and Codex's internal reviewer are left out.

claudepro lists the Claude models first and codexpro the GPT models. Within a family the newest generation comes first, and within a generation the larger line: Fable, Opus, Sonnet, Haiku for Claude; Astra, Sol, Terra, Luna for GPT.

Neither command pins a model. claudepro keeps Claude Code's own Default, Opus, Sonnet and Haiku rows, whose aliases move to the newest model with each Claude Code update, and Claude Code saves a /model choice as usual. codexpro starts on the model Codex itself would pick and keeps a /model choice too. To pin one for a run:

CLAUDEPRO_MODEL=claude-sonnet-5-5 claudepro
CODEXPRO_MODEL=gpt-6.1-sol codexpro

On Windows, set CLAUDEPRO_MODEL=... in cmd or $env:CLAUDEPRO_MODEL = '...' in PowerShell. A server owner pins a model for everyone with a model alias named claudepro-default.<model id> or codexpro-default.<model id>.

Both commands run every model, but GPT models are trained on Codex's prompt and tools and do better in codexpro. In Codex, a model the installed Codex does not ship, Claude models included, runs on Codex's gpt-5.5 settings with OpenAI-only features turned off.

Files and uninstall

Linux, macOS Windows
Commands ~/.local/bin/claudepro, codexpro %USERPROFILE%\.claudepro\bin\
Server URL and key ~/.config/claudepro/url, key %USERPROFILE%\.claudepro\url, key
Launcher files ~/.local/share/claudepro/ %USERPROFILE%\.claudepro\
Codex config and sessions ~/.local/share/claudepro/codex-home/ %USERPROFILE%\.claudepro\codex-home\

Codex keeps its own folder apart from ~/.codex, so codexpro resume lists only codexpro sessions and ~/.codex/AGENTS.md is not read.

To uninstall on Linux and macOS: rm -rf ~/.local/bin/claudepro ~/.local/bin/codexpro ~/.local/share/claudepro ~/.config/claudepro. On Windows, delete %USERPROFILE%\.claudepro and remove its bin folder from the user PATH.

About

The pro(xy) version of your coding agent.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages