Official Anthropic Documentation Links

Organized directory of official Anthropic documentation for Claude Code. Find the right reference for skills, plugins, agents, hooks, MCP, CLI usage, and best practices.

Overview

Use primary Claude Code documentation for platform behavior. This marketplace adds repository-specific metadata, validation, static grading, and contribution rules; it does not replace the platform specification or make every example a production certification.

The catalog and public search inventory change over time. Use the site’s generated public counts rather than a hard-coded plugin or skill total in this directory.

Primary reference directory

Topic Primary source What it owns
Claude Code Overview Product setup and capabilities
Plugins Plugin reference Native layout, manifest, component paths, and substitutions
Installation and management Discover plugins Marketplace registration, scopes, install, update, and uninstall
Skills and legacy commands Skills Invocation, frontmatter, context, and command-file compatibility
Subagents Subagents Agent definitions, delegation, tools, and model settings
Hooks Hooks Event schema, stdin payloads, decisions, and exit behavior
MCP integration Claude Code MCP Client configuration and supported connections
Native CLI CLI reference Claude commands and flags
Editor integrations VS Code and JetBrains Integration-specific setup and behavior

The broad /agents page covers parallel work; the specific agent-definition authority is /sub-agents. Older slash-command links may resolve to skills documentation because command files and skills share that reference.

Skills: native format versus marketplace requirements

The Agent Skills specification and Claude’s native skill reference describe their own permissive formats. This repository independently requires all eight marketplace fields: name, description, allowed-tools, version, author, license, compatibility, and tags. Missing fields are errors at marketplace tier. Do not present those stricter requirements as optional polish or silently remove them to match the native floor.

allowed-tools preapproves tools for an invocation; it is not an exclusive sandbox. Manual invocation does not require argument-hint. user-invocable: false blocks direct user invocation; disable-model-invocation separately controls model invocation.

effort is supported for skills as well as agents. Its current five values are model-dependent. Check the primary page for runtime behavior and the unchanged local schema for marketplace acceptance.

Agents and hooks

Agents support a tools allowlist and disallowedTools denylist. Do not infer native capability controls from fields named capabilities, expertise_level, or activation_priority; the local validator rejects them. Plugin agents and standalone definitions also have different local restrictions.

Hooks use named lifecycle events such as PreToolUse, PostToolUse, and SessionStart, not invented pre-tool-call keys. A command hook reads event JSON from stdin. Exit 2 is a documented blocking mechanism for applicable events; arbitrary nonzero statuses are not a universal blocking contract, and post-tool hooks cannot undo an executed action.

MCP protocol and SDK sources

MCP is not the Anthropic Messages API. @anthropic-ai/sdk is not the MCP server SDK. Use the MCP specification and the appropriate official SDK, such as the TypeScript SDK, for server development.

Select a coherent pinned SDK generation. The maintained v1 @modelcontextprotocol/sdk API and v2 APIs differ; do not combine their examples. Standard transports include stdio and Streamable HTTP; legacy HTTP+SSE compatibility is a separate concern.

This marketplace’s role

Anthropic publishes official marketplaces. Tons of Skills is a complementary catalog with its own contribution workflow, source provenance, and stricter metadata and static gates. Search results, grades, and badges do not substitute for reviewing source or verifying a live integration.

The canonical repository is jeremylongshore/tons-of-skills-marketplace. Its STANDARDS.md authority index points to the canonical local standards. This directory links to owners; it does not adopt a new schema, security policy, or kernel version.