[Claude Code Introduction E32] MCP concept and installation: Standard for bringing external tools, DB, and API into Claude Code

Оригинальный автор：새로운 시작 (neosarchizo)

Первоисточник：https://www.youtube.com/embed/jwlQYGRYsms

Many of the channel's translated audio tracks are produced with HeyGen. If you are interested in AI video translation, dubbing, and multilingual content creation, try HeyGen here: https://bit.ly/4o4bn9P You can help the channel by subscribing through this link — thank you! If you looked at Hooks that refine the control flow *inside* Claude Code in sessions 28 to 31, from session 32 it is Section 7 *MCP Server Integration*. In the opposite direction — we organize the open standard MCP (Model Context Protocol) that brings external systems such as Sentry, JIRA, PostgreSQL, and Figma into Claude Code. 📌 What we cover - MCP Core: An open standard started by Anthropic. Difference between the flow where the user *copies and pastes* data and the flow where Claude *accesses* the external system directly - three-party architecture: host (Claude Code) / client (one per server, 1:1 mapping) / server (external system). Both local and remote servers are identical in specification - two-layer separation: data layer (JSON-RPC 2.0, lifecycle·capabilities) / transport layer (stdio·HTTP). The same message flows to different transports - Three primitives and control: Tools (called by the model) / Resources (imported by the app as @, Direct·Template two types) / Prompts (called by the user with /mcp__server__name) - Comparison of the three transports: Streamable HTTP (recommended, SaaS) / SSE (deprecated) / stdio (local child process). The type field considers http and streamable-http to be the same - claude mcp add Option order: --transport·--env·--scope·--header are all before name, stdio is actual command after --. HTTP is a bearer token with --header, and OAuth is finished with /mcp - Scope set: local (default, only me) / project (.mcp.json shared with team, approval required for first use) / user (all my projects). Priority local 〉 project 〉 user - Issue of trust: Anthropic Directory is not a security audit, prompt injection risk, trust verification on first connection - Output limit and timeout: warning 10K, upper limit 25K tokens, MAX_MCP_OUTPUT_TOKENS·MCP_TIMEOUT·timeout by server - Claude Code is exposed to the MCP server with claude mcp serve (called from Claude Desktop) 📌 Environment information - Claude Code latest version (built-in MCP client) - claude mcp add / list / get / remove / add-json / add-from-claude-desktop / serve Subcommand - /mcp slash command (authentication, retry, check number of tools) - .mcp.json (project root, git commit target), ~/.claude.json (local·user scope) - Environment Variables: MAX_MCP_OUTPUT_TOKENS, MCP_TIMEOUT, CLAUDE_CODE_MCP_ALLOWLIST_ENV 📌 Chapter 0:00 Introduction — From Hook (internal) to MCP (external) 0:57 What is MCP — USB-C port for AI, meaning of open standard 1:51 Limits of three paths for AI to work without MCP 2:59 host / client / server — 3 characters Architecture and 1:1 mapping 4:13 Data layer (JSON-RPC 2.0) / transport layer separation 6:49 Three primitives — Tools / Resources / Prompts Difference in control rights 10:19 Three transmission methods — HTTP / SSE / stdio selection criteria 10:49 claude mcp add — HTTP·stdio registration, option order, authentication 13:05 Scope set — local · project · user and .mcp.json 17:43 Issues of trust — Directory·prompt injection·trust verification 19:16 Output limit·Timeout·claude mcp serve 20:51 Common pitfalls and summary 📌 Reference link - Blog: https://neosarchizo.github.io/posts/claude-code-32-mcp-concept-installation - Claude Code MCP official document: https://code.claude.com/docs/en/mcp - Model Context Protocol official site: https://modelcontextprotocol.io/ - MCP architecture (host·client·server): https://modelcontextprotocol.io/docs/learn/architecture - MCP server concept (Tools·Resources·Prompts): https://modelcontextprotocol.io/docs/learn/server-concepts - Claude Code security (MCP risk model): https://code.claude.com/docs/en/security - Managed MCP (organization-wide restrictions): https://code.claude.com/docs/en/managed-mcp - Environment variables (MCP_TIMEOUT·MAX_MCP_OUTPUT_TOKENS): https://code.claude.com/docs/en/env-vars - MCP reference server collection: https://github.com/modelcontextprotocol/servers - JSON-RPC 2.0 specification: https://www.jsonrpc.org/ #ClaudeCode #MCP #AIDevelopment