Same 18 models, same agent capabilities, zero GUI overhead. One command to install, one command to start building. Works in any shell, any OS, any CI runner.
The CLI runs wherever a terminal does: macOS, Linux, Windows, WSL, SSH sessions, inside tmux, inside screen, inside a Docker container. One command installs it globally, a second starts an interactive session. No GUI, no window manager, no extra process to keep running.
Interactive mode gives you a prompt where you type requests in plain English and get streaming responses with tool calls rendered inline. Non-interactive mode takes a single request as an argument and prints the result, making it trivial to pipe output into other commands or drop into a shell script.
Everything is keyboard-first. Slash commands swap models, flip modes, attach images, and toggle permission levels without leaving the prompt. Arrow keys cycle through history. Ctrl-C cancels a running tool call cleanly. It feels like the shell you already use, not a chat app pretending to be one.
The CLI isn't a stripped-down companion. It's the same agent runtime as the desktop app, wrapped in a terminal UI instead of an Electron window. Agent Mode, Swarm, Team, Research, and multi-model compare all work exactly the way they do in the desktop app. Code reviews, codebase indexing with Tree-sitter and BM25, semantic file search, and 3-tier permissions (observer / standard / autonomous) are all available.
Under the hood, the CLI shares the same tool definitions, prompts, and model router as the desktop app. Any improvement that ships to the desktop agent ships to the CLI in the same release. You can start a task at your desk, continue it on a remote server over SSH, and the agent's behavior is identical in both places.
Your account, your credits, and your settings are shared across surfaces. Sign in once with mulu-code login and the CLI uses the same billing, the same saved model preferences, and the same usage tracking as the desktop app. Credits are deducted from the same wallet. There's no separate CLI plan.
Type a slash, get a menu. /model switches models mid-conversation without losing context. /mode flips between Chat, Agent, and Ask. /swarm launches parallel agents on a shared task. /team runs a multi-agent collaboration. /compare fans out the same prompt to three models and shows the results side by side. /research pulls citations and writes a brief.
/permissions sets the approval level: observer (ask before every action), standard (ask for risky actions), or autonomous (run freely, log everything). /image path.png attaches an image as visual context. /autoclear toggles screen clearing between turns if you like a clean slate. Every slash command is documented inline with /help.
New commands land through normal releases. No plugin system to configure, no third-party namespace collisions. The command set is the same across the desktop app and the CLI, so you only learn it once. Tab completion makes the whole thing feel like a shell builtin rather than an escape hatch.
Headless mode runs without a TTY, which means the CLI drops into GitHub Actions, GitLab CI, CircleCI, Jenkins, and any other runner without special setup. Authenticate with an environment variable, pass your prompt as an argument, and the CLI executes non-interactively and exits with a proper status code when done.
Common CI uses include automated code review on every PR, security scans on main branch pushes, generated migration files when a schema changes, and nightly agent runs that triage open issues. Output is structured JSON on request, so downstream steps can parse tool calls, file changes, and token usage without screen-scraping.
For teams, CI runs are tied to your organization wallet with per-seat usage visibility. Every CI run shows up in the same usage dashboard as interactive sessions, so you can see at a glance how much agent budget your pipeline consumes. Rate limits, retries, and timeout behavior are all configurable with standard flags.
Built-in commands for common workflows. Type a slash and go.
/model sonnetSwitch AI model mid-conversation./mode agentEnable autonomous agent mode./swarm "build X"Launch parallel swarm agents on a task./compare "fix bug"Compare answers from multiple models./research "topic"Deep research with citations./team "build app"Multi-agent collaboration mode./permissions autoSet tool approval level./image path.pngAttach image for visual context.All 18 models available from the same binary. Switch with /model or let auto-routing choose.
One command. All the power. Works on every OS, every shell, every CI runner. Signed in once, running everywhere.