Claude Code CLI as an Agent Platform

Claude Code ships as a developer CLI, but underneath it's a general agent platform: model router, tool protocol (MCP), project memory (CLAUDE.md), subagents, hooks, skills. If you're already paying for Claude, Claude Code is the cheapest way to ship a real agent in-house.

The five extension points

Custom subagent example

---
name: accessibility-reviewer
description: Reviews web UI for WCAG 2.2 compliance.
tools: Read, Grep, Glob, WebFetch
model: claude-sonnet-4-6
---

You are a strict WCAG 2.2 accessibility reviewer.

For any files or URLs given, check:
1. Color contrast ratios (4.5:1 body, 3:1 large).
2. Semantic HTML (headings, landmarks, labels).
3. Keyboard navigation (focus order, skip links, traps).
4. Alt text on images (descriptive, not filename).
5. Form labels and error states.

Report: issues as markdown table (severity / location / fix).
Never modify files. Read-only review.

Hooks for safety

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [{
          "type": "command",
          "command": "./.claude/hooks/block-rm-rf.sh"
        }]
      }
    ],
    "Stop": [
      {
        "hooks": [{
          "type": "command",
          "command": "osascript -e 'display notification \"Claude done\" with title \"CC\"'"
        }]
      }
    ]
  }
}

MCP in Claude Code

MCP servers declared in claude_desktop_config.json (global) or .mcp.json (project) become available as tools inside every Claude Code session and to every subagent (unless you restrict via the tools frontmatter field). This is how you give an agent Supabase, GitHub, or your private internal API access.

Skills — the agent-skills primitive

Skills are folders under .claude/skills/ or ~/.claude/skills/. Each skill has a SKILL.md with a description; Claude pulls it in when the description matches the user's task. Perfect for domain knowledge ('how our billing works', 'our code style guide') that shouldn't bloat CLAUDE.md.

# .claude/skills/deploy-runbook/SKILL.md

---
description: Use this skill when the user asks to deploy, release, or
push to production. Knows our specific deploy flow and gotchas.
---

# Deploy Runbook

Our deploy is a three-step GitHub Actions workflow:

1. Tag `vX.Y.Z` on main — triggers CI.
2. CI runs tests + builds artifacts.
3. Approval gate in GitHub Environments: prod requires a second reviewer.

Never run `kubectl apply` directly. Never edit prod secrets from CLI.
Rollback: `./scripts/rollback.sh <prev-tag>`.

Common patterns

• Project-specific subagents in .claude/agents/ (code reviewer, test writer, doc updater).
• Org-wide subagents in ~/.claude/agents/ (security reviewer, deploy assistant).
• Private MCP servers for internal APIs — auth via env vars in claude_desktop_config.json.
• Hooks that post to Slack or a message queue on Stop (chain agents across sessions).
• Skills for domain knowledge (style guides, runbooks) that loads lazily.

Claude Code is underrated as an agent platform. If your team is already on Claude, skip the framework selection rabbit hole for internal automation — use the CLI.