name: ai-tools description: Manage AI CLI tools (Claude Code, Gemini CLI, Codex CLI). Use when user asks about AI tool versions, updates, installation, or wants to update their AI CLI tools. Detects installed tools, checks versions, and retrieves update instructions from authoritative sources using delegation-first pattern. allowed-tools: Bash, Skill, Task, Read, mcp__perplexity__search, WebSearch

AI Tools Management

Extensible AI CLI tools management with delegation-first design. Detects installed tools, reports versions, and retrieves update instructions from authoritative sources.

Design Principles

1. No hardcoded documentation - Never duplicate docs that exist elsewhere
2. Delegation-first - Route to ecosystem skills/agents for authoritative guidance
3. Graceful fallback - Skill → MCP research → web search
4. Extensible - Easy to add new tools via tool-registry.md
5. Smart detection - Check what's installed, report versions

Workflow

Step 1: Detect Installed Tools

Check which AI CLI tools are installed on the system:

# Claude Code
which claude 2>/dev/null || where claude 2>/dev/null || echo "NOT_FOUND"
claude --version 2>/dev/null || echo "VERSION_UNKNOWN"

# Gemini CLI
which gemini 2>/dev/null || where gemini 2>/dev/null || echo "NOT_FOUND"
gemini --version 2>/dev/null || echo "VERSION_UNKNOWN"

# Codex CLI
which codex 2>/dev/null || where codex 2>/dev/null || echo "NOT_FOUND"
codex --version 2>/dev/null || echo "VERSION_UNKNOWN"

Report which tools are installed with their current versions.

Step 2: Get Update Instructions (Delegation Pattern)

For each tool that needs updating, retrieve instructions from authoritative sources using the delegation chain:

Claude Code

Primary: Invoke claude-ecosystem:docs-management skill with query "Claude Code update install upgrade"

Fallback 1: If skill unavailable, spawn claude-code-guide agent:

First WebFetch https://code.claude.com/docs/en/claude_code_docs_map.md to find
update/install documentation. Then WebFetch those pages. Return the update commands.

Fallback 2: If agent unavailable, use MCP:

mcp__perplexity__search("Claude Code CLI update install latest version 2026")

Fallback 3: WebSearch for "Claude Code CLI update install"

Gemini CLI

Primary: Invoke google-ecosystem:gemini-cli-docs skill with query "update install upgrade npm"

Fallback 1: If skill unavailable, use MCP:

mcp__perplexity__search("Google Gemini CLI npm update install latest version 2026")

Fallback 2: WebSearch for "Google Gemini CLI npm install update"

Codex CLI

Primary: Invoke openai-ecosystem:codex-cli-docs skill with query "update install upgrade npm"

Fallback 1: If skill unavailable, use MCP:

mcp__perplexity__search("OpenAI Codex CLI npm update install latest version 2026")

Fallback 2: WebSearch for "OpenAI Codex CLI npm install update"

Step 3: Execute Updates

After retrieving authoritative update instructions:

1. Present the update commands to user
2. If --dry-run mode, stop here (show what would happen)
3. Otherwise, execute the update commands
4. Report new versions after updates complete

Step 4: Report Results

Provide a summary table:

Tool Registry

See references/tool-registry.md for the complete tool registry with:

• Detection commands
• Delegation sources (skills, agents, MCP queries)
• Fallback search queries

Adding New Tools

To add a new AI CLI tool:

1. Add entry to references/tool-registry.md with:

   • Detection command (how to check if installed)
   • Version command (how to get current version)
   • Docs source (which skill/agent has authoritative docs)
   • Search query (fallback for MCP/web search)

2. No code changes needed - the workflow automatically handles tools in the registry

Supported Modes

• Default: Update all installed tools
• --tool NAME: Update specific tool only
• --dry-run: Show what would happen without executing
• --check: Only check versions, don't update

Error Handling

• Tool not installed: Skip gracefully, report in summary
• Skill unavailable: Fall back to MCP search
• MCP unavailable: Fall back to web search
• Update fails: Report error, continue with other tools
• No authoritative source found: Report inability to update, suggest manual check

Why Delegation?

Update commands change over time:

• Package managers evolve (npm → pnpm, pip → uv)
• Installation methods change (npm global → npx)
• Flags and options get added/deprecated

By delegating to ecosystem skills that maintain current documentation, this skill stays accurate without manual updates.

Test Scenarios

Scenario 1: Check installed tools

Given: User asks "what AI tools do I have installed?" Then: Detect all installed tools and report versions

Scenario 2: Update all tools

Given: User asks "update my AI tools" Then: Detect installed tools, retrieve update instructions via delegation, execute updates

Scenario 3: Update specific tool

Given: User asks "update claude code" Then: Retrieve Claude Code update instructions via delegation, execute update

Scenario 4: Dry run mode

Given: User asks "show me how to update my AI tools without running anything" Then: Detect installed tools, retrieve update instructions, display without executing

Scenario 5: Check only mode

Given: User asks "what versions of AI tools do I have?" Then: Detect installed tools and report versions without attempting updates

Version History

Model: Claude Opus 4.5 (claude-opus-4-5-20251101) Last Updated: 2026-01-17