Omarchy AI Agents with Claude Code, Codex, and Local Models

Safely launch Claude Code, OpenAI Codex, OpenCode, and other agents in Omarchy, set a default agent, inspect usage, and run local models.

Omarchy treats coding agents as first-class desktop tools without choosing one provider for you. Several CLIs use mise-managed lazy launchers: the actual package is downloaded on first use, then authentication follows the provider's own flow.

Available agents and launch commands

The current official manual lists these entry points:

Command	Agent
`claude`	Claude Code
`codex`	OpenAI Codex
`opencode`	OpenCode
`copilot`	GitHub Copilot CLI
`agy`	Google Antigravity CLI
`crush`, `grok`, `pi`, `omp`, `ori`	Their corresponding agents or harnesses

Before first use, confirm the command resolves to Omarchy's mise stub rather than an unrelated executable earlier in PATH. Authentication, subscription limits, pricing, and data terms belong to each provider and can change independently.

Choose a default agent

Use Setup → Defaults → Agent or:

omarchy default agent <name>

After selection, Super + Shift + Ctrl + A launches the default in a dedicated terminal. A task can be passed directly:

omarchy agent prompt "Review this project"

The shell aliases a, c, cx, and cy launch the default agent, OpenCode, Claude Code, and Codex respectively. Verify aliases and behavior against your active shell and the current AI manual.

Some launchers may auto-approve actions

Omarchy's convenience entry points can run an agent in a mode that does not pause for every decision. Start in a test repository or planning mode, review workspace and credential access, and inspect the Git diff before allowing changes to important projects.

Agents Panel and usage

An Agents icon appears in the top bar after Omarchy detects coding-agent usage. The panel can aggregate subscriptions, session or weekly limits, and model-level usage for supported providers. Treat this as a convenience estimate, not the final provider invoice.

Never put API keys in screenshots, tutorials, shell history, or public dotfiles. Prefer the provider's login flow, an operating-system secret store, or a permission-controlled environment file.

The Omarchy Skill and customization

Omarchy links a system skill into supported Claude Code, Codex, Pi, Antigravity, and generic agent skill directories. It teaches agents about desktop configuration, the bar, and themes, but the official manual describes this as experimental.

Use it with a controlled sequence:

Ask for a read-only assessment and plan.
Confirm that it references current Omarchy paths.
Back up user overrides and preserve a Git diff.
Change one subsystem at a time.
Revert the change or use the official config recovery path when necessary.

Agent output is not a system source of truth. Human review is mandatory for disk, boot, account, network, and system-package actions.

Crash diagnosis

Omarchy can watch systemd-coredump and offer to hand crash evidence to the default agent. The manual entry point is:

omarchy agent crash <pid>

Core dumps and logs can expose paths, command arguments, and sensitive memory. Redact evidence before publishing it and establish reproducibility before filing upstream.

Local models and dictation

Install → AI includes LM Studio, Ollama, and dictation options. LM Studio offers a graphical model workflow; Ollama favors CLI and local-service integration. Check:

RAM or VRAM needs and quantization;
model license and commercial-use terms;
whether the local service is exposed to the network;
whether prompts, logs, or conversations leave the device.

Dictation processes microphone data, so review model downloads, recording retention, and telemetry before enabling it.

Build an agent workspace with Terminal, Tmux, and LazyVim, and prepare a system-level exit with snapshots and recovery.

Sources: Omarchy AI and Omarchy CLI.