Codeep on Omarchy

Codeep 3.9.1 or later

The Codeep CLI runs on Omarchy, the Arch Linux and Hyprland setup, like any other terminal tool. Three things fit it in particular: Codeep's colours follow the Omarchy theme, the skills Omarchy links into ~/.agents/skills load, and codeep --yolo -- "<prompt>" starts a session without stopping to ask. They arrived in 3.9.0; use 3.9.1 or later, which fixes typing and redraw problems first seen on Omarchy.

Install

Install Codeep the way Omarchy installs other CLI tools, through mise:

This only writes a small launcher, ~/.local/bin/codeep. The first time you run codeep, the launcher has mise download Codeep's Linux release binary from GitHub (x86_64 or aarch64, whichever matches your machine) and then starts it. Check the version:

Update

mise holds a new release back for a few days after it ships (its minimum release age), so a plain mise upgrade can leave you on the previous version. Omarchy's mup alias updates every mise tool, Codeep included, with that wait turned off:

To update Codeep alone:

API keys

On first run Codeep asks you to pick a provider and paste its API key. On Linux it keeps the key in the Secret Service keyring, which on Omarchy is GNOME Keyring, not in a plain-text file. Each key is its own item, labelled after the provider, for example keyring:api-key-z.ai@codeep.

If no Secret Service is available, or saving to it fails, Codeep keeps the key as plain text in its config file instead, and moves it into the keyring on a later start once the Secret Service is there.

Theme

On Omarchy, the colours Codeep paints itself (the brand red, the greys of the chat, the code colours, the warning orange and the YOLO badge) come from the current Omarchy theme. Switch the theme and an open Codeep session changes with it.

Codeep only reads ~/.local/state/omarchy/current/theme/colors.toml; it writes nothing to your Omarchy setup. A theme colour too faint to read is darkened or lightened until it reads, so light themes stay legible too.

Turning it off

Open settings and set Follow Omarchy theme to Off to keep Codeep's own colours:

The row is On by default and shown only on Omarchy. Anywhere else, Codeep's colours are unchanged.

Skills

Codeep loads skills from ~/.agents/skills, the directory agent tools share and where Omarchy links its own skills, so Omarchy's skills (such as omarchy and diagnose-crash) are there in every session. When two skills share a name, the project's .codeep/skills wins, then ~/.codeep/skills, then ~/.agents/skills. To list each skill under the directory it came from:

A skill can point the agent at files beside its SKILL.md, such as a guide; Omarchy's skills keep most of their instructions that way. Only files inside the skill's directory can be read: no .., no absolute path, and no link that leads out of it. A skill directory that is itself a link, as Omarchy's are, works. See Skill bundles for the format.

Starting Codeep

Start Codeep from a terminal, inside the project you want it to work on:

The first time in a folder, Codeep asks what access to give it and remembers the access you grant for that folder.

Starting with a prompt

For a script or keybinding that opens Codeep unattended, pass the prompt after -- and add --yolo:

This starts a new session with the prompt as its first message and does not stop to ask, for that launch only:

• Agent actions run without confirmation, and a vague task gets no clarifying questions.
• The current folder is used as the project with read & write access, so start it in the project folder, not in your home folder.
• Neither the folder access nor these confirmation settings are saved: the next plain codeep asks as before. The conversation itself is saved as a session like any other.
• The status bar shows YOLO.

Some things still ask. Writes the agent makes with its file tools to files that decide what runs later, such as .git/config, git hooks, .codeep/hooks, .codeep/skills, .agents/skills and MCP server lists, are still confirmed, and a workspace's own MCP servers still ask to be trusted the first time. A shell command the agent runs is not checked this way: under --yolo it runs without asking and can write those same files.

-p <prompt> and --prompt <prompt> do the same as -- <prompt>. Everything after -- is the prompt, and a leading / is sent as text, not run as a slash command: