Everything Claude Code can do in 2026 — MCP, subagents, hooks, and what to build

Claude Code stopped being "an AI that edits files in your terminal" a while ago. In 2026 it's an agentic runtime with a real capability stack — a protocol for wiring in external systems, isolated subagents, a lifecycle hook system, skills, and multi-agent orchestration. This is a field guide to what it can actually do, and what's worth building on top of it. Every capability below is from primary sources (Anthropic docs, modelcontextprotocol.io, the GitHub/Microsoft server repos) — linked at the end.

The five surfaces, one engine

Claude Code runs on five surfaces: the terminal CLI, the VS Code extension, the JetBrains extension, a standalone desktop app, and the web. They all share one engine — so your CLAUDE.md files, settings, and MCP servers work identically across every surface. Set something up once; it follows you everywhere.

MCP — the wiring

The Model Context Protocol (MCP) is the open standard that lets Claude Code reach external systems: data sources (local files, databases), tools (search, calculators, APIs), and workflows (specialized prompts). It's how Claude reads a Google Drive doc, updates a Jira ticket, pulls Slack data, or drives your own custom tooling.

What's worth knowing to use it well:

Four transports: remote HTTP (recommended), remote SSE (deprecated), local stdio, and remote WebSocket. HTTP supports OAuth and the --transport flag; WebSocket supports neither.
Three config scopes: local (this project, ~/.claude.json), project (team-shared via a version-controlled .mcp.json), and user (all your projects, private). Managed-settings keys (allowedMcpServers, deniedMcpServers, allowManagedMcpServersOnly, disabledMcpjsonServers) gate which servers may connect — deniedMcpServers wins, and an empty allowlist means lockdown.
OAuth 2.0 for remote servers — tokens stored securely and auto-refreshed, initiated via /mcp or (v2.1.186+) claude mcp login <name>.
Adding one is one line: claude mcp add --transport http notion https://mcp.notion.com/mcp.

Subagents — isolated context, scoped tools

A subagent runs in its own separate context window with a custom system prompt, specific tool access, and independent permissions. It does verbose work off to the side and returns only a summary — keeping your main conversation clean.

They're defined as Markdown files with YAML frontmatter in .claude/agents/ (project) or ~/.claude/agents/ (user). Frontmatter fields include name, description, tools, model, permissionMode, mcpServers, hooks, skills, memory, and isolation. The powerful bit: via mcpServers you can scope a subagent its own MCP servers — even inline definitions connected only for that subagent's lifetime — so those tool descriptions never bloat the main context.

Hooks — deterministic policy

Hooks are user-defined shell commands, HTTP endpoints, or LLM prompts that fire automatically at lifecycle points. Five handler types (command, http, mcp_tool, prompt, agent) fire on events including PreToolUse, PostToolUse, SessionStart, SessionEnd, UserPromptSubmit, Stop, SubagentStart, SubagentStop, PreCompact, PostCompact (and more).

The killer use is enforcement: a PreToolUse hook can block a tool call by exiting with code 2 or returning JSON with permissionDecision: "deny" — e.g. a block-rm.sh that greps for rm -rf and refuses it. Hooks are themselves governable via allowedHttpHookUrls, allowManagedHooksOnly, and disableAllHooks.

Skills (which now absorb slash commands)

Custom slash commands have merged into skills. A file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work identically — and your existing .claude/commands/ files keep working. Skills can inject dynamic context (including from shell commands) so /deploy can carry your actual, current deploy steps rather than a static blob.

Orchestration — teams, background agents, the SDK

Multiple agents in parallel on different parts of a task, coordinated by a lead agent.
Background agents — run several full sessions at once, watched from one screen.
The Agent SDK — build fully custom agents with your own orchestration, tool access, and permissions.

This is the layer that turns "an assistant" into "a fleet you dispatch."

The MCP servers worth connecting (2026)

The two verified standouts, both official:

GitHub's official MCP server — repos, issues/PRs, code analysis, workflow automation. If your work lives on GitHub, this is the first one to add.
Microsoft's Playwright MCP server — browser automation via the accessibility tree, so it needs no vision model; install with a single claude mcp add. This is what lets Claude reliably click, fill forms, and scrape JS-rendered pages.

Beyond those, the commonly-used reference/community servers cover filesystem, git, fetch, memory, sequential-thinking, databases (Postgres/SQLite), and search — plus domain servers (Notion, Figma, Slack, Jira). The rule of thumb: connect a server when the task genuinely needs that system, and scope credential-bearing ones tightly (project or subagent scope, OAuth where available).

What to build with it — ranked

Generic, broadly useful:

Repo/PR automation bot (GitHub MCP) — triage issues, review diffs, keep changelogs.
Browser data-pipeline (Playwright MCP + cron) — scrape a JS site on a schedule, commit results.
A custom MCP server for whatever system you own — turn an internal API into a Claude-usable tool.
Policy-enforced team setup — hooks that block dangerous commands, skills that encode your deploy/runbook steps.

And for the pattern I actually run — a solo fleet of free static sites on Cloudflare Pages, git-as-database, keyless browser-side AI, GitHub Actions cron, and a live-market-data MCP — the high-value builds are:

Live-data dashboards — a market-data MCP → daily scrape → commit JSON → static site rebuilds. (This is exactly how the oriz finance trackers work: cron fires, data lands in data/*.json, the Astro site redeploys.)
Self-notifying monitors — cron + a notifier (Telegram/ntfy), notify only on state change, dedup via the committed git-as-DB state.
Fleet-wide maintenance agents — fan out subagents to add tests, fix CI, or standardize configs across dozens of repos in one pass (what built this very blog network).
A portfolio/analysis tool on top of a connected finance MCP — allocation drift, concentration, rebalance nudges, entirely client-side and private.

The throughline: MCP gives Claude the reach (systems), subagents give it scale (parallelism), hooks give it safety (enforcement), and cron gives it autonomy (it runs without you). Compose those four and you get software that maintains itself.

Sources (primary): Claude Code overview · MCP in Claude Code · Subagents · Hooks · Settings · Slash commands / skills · Best practices · modelcontextprotocol.io. Researched via a 105-agent adversarial-verification pass; all capability claims cited above passed unanimous (3-0) verification against these primary sources. Version-gated details (e.g. claude mcp login from v2.1.186) track the 2026 docs — re-check against your installed version.