omnirush
v3.0.8
Published
Omnirush — free daily tokens for the most powerful coding model on earth.
Maintainers
Readme
omnirush gives you free access to frontier coding models in your terminal. GPT 6 Astra is the default; GPT 6.1 Sol and the other models in your account's catalog are one flag away.
Install
Requires Node.js 22.19 or newer on macOS, Linux or Windows.
npm install -g omnirushThe omnirush package has no dependencies and no install scripts. npm also
installs one @omnirush-ai/cli-<os>-<cpu> package for your platform with the
prebuilt runtime, search tools and engine binaries.
Quick start
omnirush login # sign in at omnirush.ai through your browser
cd my-project
omnirush # start an interactive sessionomnirush -p "<prompt>" answers once and exits: the reply goes to stdout,
errors to stderr, and the exit code is 1 when the turn fails. Piped input is
read too (git diff | omnirush -p "review this"), and a run without a
terminal (a script, CI) answers once the same way.
Features
One engine. Every run uses the omnirush engine: the full-screen interface in a terminal, and
-p, piped input, scripts and CI without it. The classic engine has been retired;--engineis no longer needed.Model and effort control. Pick a model with
-m <model>and a reasoning effort withctrl+t,--thinking <level>or-m <model>:<effort>(levels fromminimaltomax; sessions start onxhigh).Resumable sessions.
--continueresumes your last session;-s <session>reopens a specific one (the command is printed when a session ends),-rpicks one from the list. Chats from the classic interface open here too.Asks before risky actions. Three permission modes, switched with
/permissionsin the app oromnirush permissions <mode>:- Normal (default) asks only before risky actions: recursive deletes
outside temp folders,
sudo,git push/ force pushes / hard resets, changes or reads outside the project, reading secrets (.env, SSH keys, credential files), piping a download straight into a shell and publishing packages. - Careful asks before every file change and every command; reading and searching run freely.
- Auto never asks.
A question takes one key:
yallow once,aallow always (until you close omnirush),nreject. Keys you type while a question appears go to your draft, not to the question; pressEscwhen you want to answer.omnirush -pand runs without a terminal use Auto unless--permissions careful|normalis given; there a question nobody can answer is declined and the run says so on stderr. A project's ownopencode.jsonpermission rules still apply on top of the mode. Normal is a guard rail, not a sandbox: it matches command text, so a script or one-liner that deletes or pushes from inside (python -c …,node -e …),chmod/kill, and the agent's search tools over a.envfile are not asked about; use Careful (or the sandbox) when that matters.- Normal (default) asks only before risky actions: recursive deletes
outside temp folders,
Tells you when it is done. When a task finishes, or waits for your approval, while you look away, the terminal rings its bell and shows a system notification (OSC 9: iTerm2, WezTerm, Windows Terminal, kitty, Ghostty and others).
omnirush notify bell|osc9|offchanges it;omnirush notify testshows what your terminal does.Sub-agents and background jobs. Delegate work to sub-agents (
/subagents,/agents) and keep long shell commands running in the background (/jobs).MCP servers. Local MCP servers from a supported list (filesystem, git, SQLite and others; the list comes from the server) start from your
opencode.json; remote servers and servers that reach outside the project are blocked with the reason.omnirush mcpshows this folder's servers; see docs/mcp.md.OmniRush agent prompt. The agent works from omnirush.ai's own instructions, as in the desktop app, instead of the engine's generic ones.
Best practices. Built-in coding guidance, plus guides for setup, features, bugs, tests, refactors, builds, reviews and handoffs that the agent reads when a task needs them. On by default; turn it off with
"bestPractices": falsein~/.omnirush/settings.json, or for one run withOMNIRUSH_BEST_PRACTICES=0.Voice input.
/voicelets you dictate prompts (hold Space to talk).Good session guide.
/goodshows an on-demand checklist for the current session;omnirush goodexplains how Good sessions ★ earn rewards. Quitting while a turn runs asks first.Files used. OmniRush also saves the files the agent reads, runs or creates during a session (including temporary files and a few allowlisted config files, with secrets removed), so the session can be replayed.
omnirush files-used statusshows your account's state.Keeps your computer awake while the agent works. From the moment you send a prompt until its turn ends (or is stopped, or fails), and while the uploads finish after you quit, the computer does not go to sleep. An idle chat never keeps it awake, and the display still turns off and locks as usual. It uses
caffeinateon macOS, a logind inhibitor (systemd-inhibit) on Linux and a small helper on Windows; each one ends with omnirush, even after a crash. Nothing needs admin rights and no power setting is changed. Over SSH, in CI, in containers and in WSL it is skipped. On by default; turn it off withomnirush keep-awake off(saved as"keepAwake": falsein~/.omnirush/settings.json), or for one run withOMNIRUSH_KEEP_AWAKE=0.omnirush keep-awake statusshows whether it works on this machine.Headless Blender edits.
omnirush blenderapplies an edit script to a saved.blendfile and writes the result to a new file.
Commands
| Command | Description |
| --- | --- |
| omnirush | Start an interactive session in the current folder |
| omnirush -p "<prompt>" | Run a one-shot prompt and exit |
| omnirush login | Sign in through your browser |
| omnirush logout | Remove the stored sign-in tokens |
| omnirush whoami | Show the signed-in account |
| omnirush doctor | Check sign-in, runtime and agent configuration |
| omnirush update | Check for and download the latest release now |
| omnirush good | How to make a Good session ★ |
| omnirush files-used status | Show whether the files the agent uses are saved with your sessions |
| omnirush sandbox [on\|off\|status\|setup] | Run this folder in a Linux sandbox, or check it |
| omnirush keep-awake [on\|off\|status] | Keep the computer awake while the agent works (on by default) |
| omnirush permissions [careful\|normal\|auto] | When the agent asks first (Normal by default) |
| omnirush notify [auto\|bell\|osc9\|off\|test] | Bell / system notification when a task is done or needs approval |
| omnirush port | In WSL, start a new chat from a summary of a Windows chat |
| omnirush sessions import <path> | Copy CLI sessions into the current project |
| omnirush blender <file.blend> --script <edit.py> | Edit a saved Blender project headlessly |
| omnirush --version | Print the installed version |
Common options
| Option | Description |
| --- | --- |
| -m, --model <model[:effort]> | Model for this run, with optional effort (for example gpt-6-sol:high) |
| --thinking <level> | Reasoning effort for this run |
| --continue | Resume the last session |
| -s, --session <id> | Resume a specific session |
| -p, --print | Answer once and exit (also when input is piped or there is no terminal) |
| --permissions <mode> | careful, normal or auto for this run (--yolo is auto) |
| -i, --image <file> | Attach an image to the first prompt (repeatable) |
Run omnirush --help for the full list of options, session commands and
environment variables.
Sandbox
A project folder can run in a Linux sandbox: the agent's commands run inside
it, while the app, your chats and your files stay on your machine. Packages
installed in the sandbox are still there next time. Depending on your system
and OmniRush's settings, new chats may use it by default;
omnirush sandbox status shows the setting in force and which tool is used.
omnirush sandbox on # this folder runs in the sandbox from now on
omnirush --sandbox # or just this run
omnirush sandbox off # run this folder directly on your machine
omnirush sandbox setup # guided setup when a sandbox tool is missingIt uses the tool your computer has: bubblewrap, Podman or Docker on Linux
(bubblewrap ships in the Linux packages; on Ubuntu 23.10 and newer run
sudo omnirush sandbox enable once), and container tools on macOS and
Windows, including OmniRush's own WSL image. On macOS and Windows, a project
that already has builds for your machine keeps running on your machine. When
no sandbox tool is available, the run carries on directly on your machine
with a one-line note. Without the sandbox, the agent runs with your normal
local tools and permissions.
Updating
New releases are downloaded in the background and applied once running sessions have finished. To update immediately, or to remove the CLI:
omnirush update
npm uninstall -g omnirushSet OMNIRUSH_AUTO_UPDATE=0 to disable automatic updates. Sign-in tokens,
settings and sessions are stored in ~/.omnirush. See
docs/update-and-uninstall.md for details.
Good sessions ★
Good sessions earn the gold wheel (omnirush spin). How to make a Good
session:
- Work inside your project folder.
- Build or change real code.
- Run it or its tests.
- Let the last turn finish.
Native Windows, macOS, Linux and WSL all count the same. Sub-agents are not needed. One long turn (20+ min of agent work) or two real turns count. The checklist is a hint; the server's check decides.
Windows
OmniRush runs natively on Windows, and native Windows sessions count the same
as macOS, Linux and WSL sessions. WSL is optional. Chats do not move between
Windows and WSL; if you do switch, omnirush port in WSL starts a new chat
from a short handoff of a Windows chat. See
docs/session-migration.md.
After npm install -g omnirush, open a new terminal window: a terminal
that was open before the install keeps its old PATH and answers "'omnirush'
is not recognized". If a new terminal still does not find it:
npx omnirush doctornames the cause and the fix (thelauncherline).- npm's folder (
npm config get prefix, usually%APPDATA%\npm) must be on your PATH. Add it once in PowerShell, then open a new terminal:[Environment]::SetEnvironmentVariable("Path", [Environment]::GetEnvironmentVariable("Path", "User") + ";$env:APPDATA\npm", "User") - If PowerShell says "running scripts is disabled on this system", run
omnirush.cmd(ornpm.cmd), or allow local scripts once:Set-ExecutionPolicy -Scope CurrentUser RemoteSigned. - Until then,
npx omnirushstarts it too.
Development
npm testTests use the Node.js built-in test runner. Release notes are in CHANGELOG.md.
License
omnirush is licensed under the Apache License 2.0.
The omnirush engine is an omnirush-branded build of
opencode, used under the MIT License; its
license text ships in the platform packages as LICENSE-opencode-MIT. Other
bundled components keep their own licenses, listed in
THIRD_PARTY_NOTICES.md.
Contact: [email protected]
