Skip to main content
Version: v0.1.2

mekara install

Install mekara components: MCP hooks, bundled commands, or both.

Usage​

mekara install           # Install both hooks and commands
mekara install hooks     # Install MCP hooks only
mekara install commands  # Install bundled commands only

Subcommands​

mekara install (no subcommand)​

Installs both hooks and commands. This is the recommended way to set up mekara for the first time.

mekara install hooks​

Sets up MCP server and hook integration by creating or updating:

  • ~/.claude.json — Declares the mekara MCP server to Claude Code
  • ~/.claude/settings.json — Configures hooks (UserPromptSubmit and PreToolUse) and auto-allows mekara MCP tools
  • ~/.config/opencode/opencode.json — Configures the mekara MCP server for OpenCode

Error handling:

  • If a config file contains invalid JSON (e.g., comments, which JSON doesn't support), that file is skipped with a warning
  • If the hooks installation fails for any reason, a fallback message suggests running /ai-tooling:setup-mekara-mcp in Claude Code to complete setup interactively

mekara install commands​

Installs bundled components to ~/.agents/:

  • Skills → ~/.agents/skills/ (both NL and compiled together)
  • Standards → ~/.agents/standards/

Also sets up a symlink from ~/.claude/skills/ to ~/.agents/skills/, making mekara's bundled commands available globally.

Why both NL and compiled are installed

The skill resolution algorithm finds both NL sources (SKILL.md) and compiled versions (mekara.py) in unified skill folders. For bundled skills to be available when user-level skills exist, both versions are installed together to the same location.

Standards installation:

Standards are installed to ~/.agents/standards/. Commands that reference standards using @standard:name syntax have those references replaced with actual file paths (e.g., @~/.agents/standards/command.md) so Claude Code's @ file reference mechanism can resolve them.

Symlink behavior:

  • ~/.agents/skills/ is the canonical directory
  • ~/.claude/skills/ is a symlink to ~/.agents/skills/

This matches the standard project convention where the generic Agent Skills directory is canonical.

File handling:

  • Preserves directory structure (e.g., project/setup-docs.md)
  • Skips files that already have the same content
  • Updates files that have changed
  • Transforms @standard:name references to absolute file paths

When to Use​

Run mekara install when you want to:

  • Enable mekara script execution via /command-name in Claude Code
  • Get access to mekara's bundled commands globally
  • Set up a new machine to use mekara

Run mekara install hooks when you only need to update hook configuration without touching commands.

Run mekara install commands when you want to update the bundled commands to the latest version.

Example​

mekara install

# Output:
# Setting up mekara MCP integration...
# Done.
# Created symlink: /Users/you/.claude/skills -> /Users/you/.agents/skills
# Installed 35 skills to /Users/you/.agents/skills
# Installed 8 standards to /Users/you/.agents/standards

After installation, restart Claude Code (or reload the project) and you can run mekara scripts by typing /command-name in the chat.

Installed Components​

MCP Server​

See mekara mcp for details about the MCP server and the tools it provides.

Hooks​

The installation configures three hooks that enhance the mekara integration:

UserPromptSubmit Hook​

Intercepts user-typed commands like /finish or /test:random:

  • Detects commands starting with /
  • For compiled scripts: outputs instructions telling Claude to use the MCP tools
  • For bundled natural-language commands: injects the command content directly into the conversation

PreToolUse Hook (Skill Redirection)​

Prevents Claude from using the Skill tool for compiled mekara scripts:

  • When Claude tries to invoke a compiled script via Skill, blocks it and redirects to MCP
  • Essential for nested script invocations to work correctly

PreToolUse Hook (Auto-Approve)​

Reduces friction from Claude Code permissions bugs by auto-approving most operations:

  • Auto-approves all tool calls except bash commands starting with rm or git commit
  • Those dangerous commands still require user confirmation
  • Note: This is a minimal workaround for upstream permissions bugs. For more comprehensive permissions control, see claude-code-permissions-hook

See Also​