Skip to content
CLIOSearchCtrl K
Download

Commands and @ mentions

Type / in the message box to run a command, or @ to bring a subagent, a workflow, a flow or a tool server into your message. Both menus list only what this conversation’s agent has.

Run a command

Section titled “Run a command”

  1. Type /. The menu groups what you can run: Commands, then Skills, then each tool server’s prompts.

  2. Keep typing to narrow it, use the arrow keys to choose, and press Enter. Words after the command are its arguments, for example /component Button primary.

Each row says where it comes from (a plugin, this workspace or your own folder) and what it takes. A row that cannot run says why in words: Reserved name, Shadowed or Unavailable. Hover the label for the reason.

Commands, skills and prompts

Section titled “Commands, skills and prompts”
  • Commands run as a message to the agent: the command file’s text, with your arguments filled in, becomes what you sent. The built-in ones (/clear, /goal, /loop, /cron and its alias /schedule, /cache-stats and /dump-trace) do their job at once.
  • Skills load a skill for the agent before it reads your message, so it works from that skill’s procedure. Your message records that you loaded it. A skill marked user-invocable: false is not listed; the agent can still load it when it needs it.
  • Prompts come from a tool server the agent is connected to, named /mcp__<server>__<prompt>. Your words fill the prompt’s arguments in order, the last argument taking the rest. If a required argument is missing, the command says which one. The fetched prompt becomes your message, with any image or document it includes; open Prompt … from … under your message to read it. When a server adds a prompt while you work, it appears in the menu the next time you type /.
The prompt's text, the resource it named and the assistant message it included, opened under the message that sent it.

When two commands share a name

Section titled “When two commands share a name”

Two plugins can ship a command of the same name. The menu then lists both under their full names, such as /commands-fixture:review-notes and /commands-clash:review-notes, marked Same name elsewhere. Type the full name to run one; the short name alone is refused and the refusal names both.

Write your own commands

Section titled “Write your own commands”

A command is a Markdown file. Its name is the file name, and its text is what the agent receives:

.clio/commands/standup.md
---
description: Summarize what changed since yesterday.
argument-hint: "<what you did>"
---
Write a three-line standup: yesterday, today, blockers. Yesterday: $ARGUMENTS
Where the file goes Who sees it Full name
.clio/commands/ in the workspace this workspace /project:<name>
.claude/commands/ in the workspace this workspace /project:<name>
~/.config/clio-agent/commands/ you, in every workspace /user:<name>
~/.claude/commands/ you, in every workspace /user:<name>
a plugin’s commands/ folder conversations whose agent has the plugin /<plugin>:<name>
  • $ARGUMENTS is everything typed after the command; $1, $2 and so on are its words, and quotes keep words together.
  • A file in a subfolder keeps its name; the folder becomes its label in the menu.
  • When the same name exists in several of these folders, the first in the table wins and the others are listed as Shadowed, naming the file that wins.
  • A built-in name, such as goal, is reserved: a file named goal.md is listed as Reserved name and /goal still runs the built-in.
  • agent: <subagent> or workflow: <workflow> in the frontmatter asks for that subagent or workflow, the same as writing @<name> in the message.
  • model and allowed-tools from Claude Code commands are read but not used. The menu says Some fields ignored.

Bring in a subagent, a workflow, a flow or a tool server with @

Section titled “Bring in a subagent, a workflow, a flow or a tool server with @”

  1. Type @. Your workspace’s files are listed; Agents (the subagents this agent can start), Workflows (with the inputs they take), Flows and Tools (its tool servers) are collapsed, each with its count. Open one to browse it.

  2. Keep typing: every category shows what matches. Press Enter, and the name becomes a chip in your message. You can add several.

  3. Send the message.
A few letters show the matches in every category.
  • A subagent, workflow or flow you name is what you ask the agent to use for this message. Nothing starts before the agent reads it: the agent writes the brief from the whole conversation (so “the folder we talked about” becomes the folder’s path), sets what the work needs (a model, a workflow’s inputs, a flow’s options), and starts it. Under your message you see You asked for explore on this message; under the answer, Started explore as you asked, and the work appears with its progress like any other, started by you. If the agent did not start it, the answer says so: explore was not started, with the reason when a call was refused. The agent keeps every tool while it works.
  • A tool server you name has its tools loaded for this message, so the agent can use them straight away. Under your message: You pointed this message at the hdf5 tools.
The agent wrote explore's brief from the conversation, waited for its report and used it.
Nothing is forced: when the agent does not start what you asked for, the answer says so.
The server's tools are loaded for the message, and the agent calls them without searching for them first.
  • Typing the full name also works: @explore as its own word becomes a chip when you send. An address such as bob@example.org stays text.
  • A name the agent has no subagent or server for never becomes a chip; it is sent as ordinary text.

Plugins written for Claude Code

Section titled “Plugins written for Claude Code”

CLIO installs a Claude Code plugin as it is, through a translation: an importer reads its files into CLIO’s own plugin records, and a hook adapter runs its hooks on the same wire Claude Code uses. clio-agent plugin check names anything that has no CLIO counterpart instead of dropping it silently. The writing method in Write a paper is one such package, brought into CLIO.

In the Claude Code plugin In CLIO How it maps
commands/*.md / commands, /<plugin>:<name> exactly, with $ARGUMENTS and $1, $2
skills/*/SKILL.md skills, also / commands exactly
agents/*.md subagents exactly; each tool it names maps by the table below
hooks/hooks.json hooks on the same events: PreToolUse, PostToolUse, UserPromptSubmit, UserPromptExpansion, Stop, SubagentStop, SessionStart, SessionEnd, PreCompact exactly; another event is reported as unmapped
.mcp.json tool servers exactly, with ${CLAUDE_PLUGIN_ROOT}
Read, Glob, Edit, MultiEdit, Write the workspace file tools exactly
Bash the shell tool exactly
WebFetch, WebSearch the web tools exactly
Task, Agent spawn_agent_task exactly
Grep the workspace text search in part: its regex, glob and files, content and count modes map; multiline mode and split -A/-B context do not, and plugin check says so
another tool a subagent names left out of that subagent plugin check warns
model, allowed-tools, disable-model-invocation in a command read, not used the menu says Some fields ignored
output-styles/, outputStyles, lspServers not used plugin check warns
Section titled “Related”