asepharyana
ec3a2867d4
- bun workspaces + Turborepo monorepo (apps/web, apps/mcp; packages/*, scripts) - @mcpedia/core single business-logic layer (Document/Content/Search services) - @mcpedia/db Drizzle schema: documents + weighted tsvector (GIN) for FTS - @mcpedia/parser frontmatter, @mcpedia/search Postgres FTS (ts_rank+ts_headline) - Next.js 16 Web UI (home/doc SSG, search dynamic) + react-markdown render - MCP server (stdio) with 4 tools + in-memory smoke test - scripts/indexer walks content/ -> upserts into Postgres - 4 seed docs; README + PHASES status
35 lines
916 B
Markdown
35 lines
916 B
Markdown
---
|
|
id: mcp-architecture
|
|
title: MCP Architecture Notes
|
|
type: research
|
|
tags:
|
|
- mcp
|
|
- architecture
|
|
- ai
|
|
status: published
|
|
author: asep
|
|
created_at: 2026-08-19
|
|
updated_at: 2026-08-19
|
|
---
|
|
|
|
# MCP Architecture Notes
|
|
|
|
The Model Context Protocol (MCP) lets an AI client treat a knowledge base as a
|
|
first-class context source instead of yet another REST API. The server exposes
|
|
tools and resources; the client decides what to read.
|
|
|
|
## Tools vs resources
|
|
|
|
- Tools are actions the model calls (`search_documents`, `get_document`).
|
|
- Resources are addressable content the model can pull (`mcpedia://docs/...`).
|
|
|
|
## Why it matters here
|
|
|
|
MCPedia exposes both. The Web UI is for humans; the MCP server is for agents.
|
|
Both go through the same Core layer, so there is exactly one copy of the
|
|
business logic and one search implementation.
|
|
|
|
## Reference
|
|
|
|
Related reading: the WebSocket contract and the tRPC type-safe API notes.
|