Cc MCP

MCP

Model Context Protocol servers extend Claude Code with tools, resources, and prompts. This skill covers selection, configuration, and the three server primitives.

The three primitives

Cost model: historically every tool's name + description + JSON schema consumed context every turn. As of 2026, Claude Code defers tool schemas by default and discovers them via ToolSearch (see below), so the per-turn drain is mostly limited to built-ins and alwaysLoad servers. Still prefer MCP Prompts/Resources for heavy reference material Claude loads only when asked.

Configuration files

Project-scoped is almost always better — avoids global context cost when you're not in that project.

Transports

Declare the transport explicitly with "type": "http" | "stdio" | "sse" in the server entry. HTTP/SSE servers use url (+ optional headers/oauth); stdio servers use command/args/env.

Tool Search & deferred tools

As of 2026, Claude Code defers MCP tool schemas by default instead of loading every tool's name + description + JSON schema into context up front. Claude discovers tools on demand via the built-in ToolSearch tool, then calls them normally. This is what lets a session connect to dozens of MCP servers (thousands of tools) without drowning the context window.

• Default behavior: a connecting server's tools appear by name only; their schemas load when ToolSearch matches them to the task.
• Force a server's tools to always load (skip deferral): set "alwaysLoad": true on the server entry — use only for small, always-needed servers.
• Disable globally: ENABLE_TOOL_SEARCH=false (rarely worth it — you trade context for eager loading). With deferral off, WaitForMcpServers is available to block until background servers finish connecting.
• Output caps: large tool results are truncated at a default token ceiling; raise it per session with MAX_MCP_OUTPUT_TOKENS.

Implication for this plugin's design: the old "every tool costs context every turn" math is now mostly paid only for alwaysLoad servers and built-ins. Still prefer MCP Prompts/Resources for heavy reference material, but the deferral default means a few extra servers are no longer the liability they once were.

Top recommendations (every project)

Recommendations by stack

Use cc_docs_hook_pack_recommend / cc_docs_team_topology_recommend style logic:

Obsidian MCP — first-class integration

The user's vault is at C:/Users/MarkusAhling/obsidian/. If the Local REST API plugin is installed in Obsidian, expose it via:

{
  "mcpServers": {
    "obsidian": {
      "command": "npx",
      "args": ["-y", "obsidian-mcp-server"],
      "env": { "OBSIDIAN_API_URL": "http://127.0.0.1:27123", "OBSIDIAN_API_KEY": "..." }
    }
  }
}

Claude then reads/writes vault notes via mcp__obsidian__* tools — used extensively by the memory-consolidator (tier 2 writes).

Channels (event-driven MCPs)

Channels are MCP servers that receive external events (webhooks, messages) and expose them to Claude. Four patterns:

Fetch implementation via cc_kb_channel_server(name) — returns full TypeScript source.

Debugging connection issues

1. Server doesn't start: check stdin/stdout isn't polluted by logging. MCP servers must write logs to stderr only.
2. Server crashes silently: run the command manually (npx -y server-name) to see stderr.
3. Tools not visible to Claude: check capabilities.tools is declared in server init, and ListToolsRequestSchema handler exists.
4. Slow tool calls: check for synchronous file I/O or network calls in hot paths.

MCP delegation

Anti-patterns

• 20+ MCP servers globally → 50k+ tokens passive context cost.
• Putting secrets in .mcp.json → commits leak keys. Use env vars.
• stdin logging in an MCP server → breaks the protocol framing.
• Same MCP server installed globally AND project-scoped → duplicate tool definitions.
• Skipping HMAC verification on webhook channels → anyone can spam your Claude.