Spokes.wiki Search About
Defined Term standard updated Thu Jun 18 2026 00:00:00 GMT+0000 (Coordinated Universal Time)

Model Context Protocol (MCP)

An open protocol (modelcontextprotocol.io) for wiring an LLM to external data and tools via standardized MCP servers. The spec’s own framing is “a USB-C port for AI applications” — one standardized connector replacing bespoke per-integration glue (mcp-spec-introduction). It is a bridge node: it connects this wiki’s knowledge-management cluster (A) with the agent tooling now in the sibling agentic-tooling-wiki (B) — which is why it stays here and is linked cross-wiki.

How it actually works

From the protocol’s own architecture overview (mcp-spec-introduction):

  • Client-server, one client per server. An MCP host (the AI app — Claude Code, Claude Desktop, VS Code) spins up one MCP client per MCP server; each client holds a dedicated connection. The server “provides context to MCP clients,” local or remote.
  • Two layers. A data layer on JSON-RPC 2.0 (lifecycle + capability negotiation
    • the primitives) over a transport layer — either stdio (local, no network overhead) or Streamable HTTP (POST + optional Server-Sent Events for remote servers; OAuth recommended). This is exactly the stdio + HTTP/OAuth surface gbrain exposes.
  • Server primitives: Tools (functions the model can invoke), Resources (context data), Prompts (reusable templates) — each discoverable via */list.
  • Client primitives (reverse direction): Sampling (a server asks the host’s LLM for a completion, so the server can stay model-independent), Elicitation (ask the user), and Logging. The handshake negotiates which primitives each side supports before any work, and servers can push list_changed notifications to keep a client’s tool registry current.

Where it shows up

  • In the agent tooling (agentic-tooling-wiki): claude-financial-services uses MCP to reach 10+ data providers (Daloopa, Morningstar, FactSet, Moody’s, LSEG, …); claude-skills-ppc frames skills × MCP = agency (live Google Ads data + the ability to execute changes).
  • In the knowledge-management cluster (here), qmd ships as an MCP server so an LLM can use it as a native search tool over a markdown wiki.
  • gbrain exposes 30+ tools over MCP (stdio + HTTP with OAuth 2.1) to a wide range of clients (Claude Code, Cowork, Cursor, ChatGPT, Perplexity) — making it the clearest example in the wiki of MCP as the universal agent-to-knowledge interface.

Why it matters here

MCP is the common substrate beneath both clusters: the standard way an LLM agent gets access to external knowledge/tools. For the llm-wiki pattern specifically, an MCP-exposed search tool (qmd) is the recommended upgrade path once the index-file approach stops scaling.

qmd · gbrain · claude-financial-services · claude-managed-agents

Related pages

Links to

Linked from