DevNotes MCP
A notes server that demonstrates advanced MCP features like sampling, elicitation, roots, pagination, and OAuth 2.1, enabling users to manage in-memory notes via natural language through a client.
README
<div align="center">
🗒️ DevNotes MCP
A notes server built to exercise every advanced feature of the Model Context Protocol.
Follow-up project for Anthropic Academy's Model Context Protocol: Advanced Topics course.
<img src="assets/certificate.png" alt="Certificate of completion — Model Context Protocol: Advanced Topics — Lazaro Gomez Vitolo" width="600">
</div>
🎓 What this is
DevNotes is a small, in-memory notes server — think a minimal Notion, exposed over MCP instead of a web UI. It's deliberately not trying to be a real product; every tool exists to be a clean, runnable example of one specific advanced MCP feature, and the notes themselves are just the excuse to have something worth calling a tool on.
| Course topic | Where it lives |
|---|---|
| Sampling — server asks the client's LLM to do the work | summarize-note |
| Elicitation — server pauses to ask the user something | create-note (priority, if omitted) |
| Roots — client grants the server a specific local folder | index-roots |
| Progress & logging notifications | also index-roots, via sendNotification/sendLoggingMessage |
| Pagination | search-notes, cursor-based |
| Structured output | every tool returns structuredContent alongside its text |
| stdio vs. Streamable HTTP transports | stdio.ts / http.ts |
| Resumable transport | eventStore.ts, stateful HTTP mode |
| Stateless transport | http.ts --stateless |
| OAuth 2.1 + PKCE authorization | auth.ts, http.ts --oauth |
🏗️ Architecture
flowchart LR
subgraph client["interactiveClient.ts"]
cap["declares: sampling, roots, elicitation"]
end
subgraph transport["pick one"]
stdio["stdio.ts"]
http["http.ts (Streamable HTTP)"]
end
subgraph server["createServer.ts"]
t1["create-note<br/><i>elicitation</i>"]
t2["search-notes<br/><i>pagination</i>"]
t3["summarize-note<br/><i>sampling</i>"]
t4["index-roots<br/><i>roots + notifications</i>"]
store[("NotesStore<br/>(in-memory)")]
end
auth["auth.ts<br/>OAuth 2.1 + PKCE"]
evt["eventStore.ts<br/>resumability"]
client <-->|JSON-RPC over stdio or HTTP| transport
transport --> server
t1 & t2 & t3 & t4 --> store
http -. "--oauth protects" .-> auth
http -. "stateful mode uses" .-> evt
Sampling is the one genuinely bidirectional flow in the demo — the server calls back into whichever process holds the model credentials — so it's worth seeing on its own:
sequenceDiagram
participant You
participant Client as interactiveClient.ts
participant Server as DevNotes server
You->>Client: summarize-note
Client->>Server: tools/call
Server->>Client: sampling/createMessage
alt ANTHROPIC_API_KEY set
Client->>Anthropic API: messages.create()
Anthropic API-->>Client: completion
else no key configured
Client->>Client: return a mock completion
end
Client-->>Server: CreateMessageResult
Server-->>Client: tool result (the summary)
Client-->>You: print it
The server never holds a model API key itself — that's the point of sampling. Only the client in this diagram needs one, and only optionally (see Quick start).
🚀 Quick start
npm install
cp .env.example .env # optional — see below
Everything runs with zero configuration. .env only matters if you want real Claude
completions for summarize-note instead of a mocked one — set ANTHROPIC_API_KEY there.
Local (stdio) — what Claude Desktop and the MCP Inspector use:
npm run client # spawns the server itself, opens an interactive prompt
# or, to poke at it directly:
npm run inspector
Remote (Streamable HTTP):
npm run start:http # stateful, resumable — default
npm run start:http:stateless # no session id, horizontally scalable
npm run start:http:oauth # requires a bearer token (see below)
npm run start:http:oauth-strict # + RFC 8707 resource-indicator enforcement
npm run client:http # talk to whichever one is running on :3000
Try roots by pointing the client at the bundled example notes:
npm run client -- --root ./examples
# then call index-roots from the prompt
Try the full OAuth flow: run npm run start:http:oauth, then open the MCP Inspector
(npx @modelcontextprotocol/inspector) and connect to http://127.0.0.1:3000/mcp — its
built-in "Connect" flow walks through dynamic client registration, the authorization
redirect, and the PKCE token exchange for you. auth.ts has a header-comment listing
exactly what it does and doesn't do — it's a real OAuth 2.1 authorization server, but an
intentionally minimal one (in-memory storage, no real login screen); the same file explains
what a production deployment needs to add back.
📁 Project structure
src/
server/
createServer.ts factory: builds one McpServer + registers all four tools
store.ts the in-memory NotesStore (create/search/importFromRoot)
stdio.ts entrypoint: stdio transport
http.ts entrypoint: Streamable HTTP, stateful/stateless/oauth
auth.ts demo OAuth 2.1 + PKCE authorization server
eventStore.ts in-memory EventStore for resumable HTTP
tools/
createNote.ts elicitation
searchNotes.ts pagination
summarizeNote.ts sampling
indexRoots.ts roots + progress/logging notifications
client/
interactiveClient.ts demo client: implements the handlers a server can call back into
examples/ sample notes for the index-roots demo
assets/ the certificate this repo is a follow-up to
🔒 Notes on the security-relevant bits
- Roots are enforced entirely by the client —
store.importFromRootonly ever reads thefile://URI it's handed, one directory level deep, capped at 200 files / 200 KB each. The server never picks its own path. createMcpExpressAppenables DNS-rebinding protection automatically for127.0.0.1— seehttp.ts.auth.tsis explicitly a demo authorization server (see the header comment in that file for the exact list of what it skips). In a real deployment you'd almost always pointverifyAccessTokenat an existing IdP rather than writing one.- Every server-initiated request (
createMessage,listRoots,elicitInput) passesrelatedRequestId, so responses route to the right session under concurrent HTTP load instead of the first one that happens to be listening.
🗺️ Learning path
Part of a series built while working through Anthropic Academy:
claude-api-starter— Building with the Claude APImcp-client-starter— Introduction to Model Context Protocolmcp-advanced-starter— MCP: Advanced Topics (this repo)- Claude Code in Action — next up
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。