Model router
o4 connects to more than 13 model providers. You can change the model during a session. If a provider fails, the router moves the request to your fallback provider.
o4 is an AI coding agent for your terminal.
It runs on your computers. It connects only to the model providers that you configure.
~$ curl -fsSL https://install.open4rena.ai/o4 | bashClick to copy the command. The installer examines the checksum before it installs o4.
macOS (Apple Silicon) · Linux (x86_64, static binary)
o4 connects to more than 13 model providers. You can change the model during a session. If a provider fails, the router moves the request to your fallback provider.
o4 makes a checkpoint at each turn. Use /rewind to move the conversation and the files back together. You can fork a checkpoint to try a different path.
Give one objective with /campaign. o4 divides it into tasks, runs worker agents in isolated worktrees, and validates each result before it continues.
o4 can plan the work before it changes files. Use /goal to set an objective. o4 keeps that objective in view across many turns.
Rules allow, ask, or deny each tool pattern. Commands run in a sandbox. Hooks can examine, change, or stop each tool operation before it runs.
o4 keeps a project-scoped memory across sessions with /remember and /recall. A tree-sitter index shows it the symbols and dependencies of your codebase.
o4 can edit files and show a diff for each change. It can run commands, search code, run tests, and use LSP servers and formatters.
You can close the terminal during a task. Use /resume to continue the session on the same machine, exactly where it stopped.
You do not need an account to install or use o4. Model traffic goes only to the providers that you configure. Nothing is proxied through our servers.
In development
The Arena is an SSH workspace for developers, built by developers. You log in to your workspace through SSH. Your o4 sessions run in the cloud and stay available from any machine. The Arena is in development.
ssh open4rena.ai