Inspiration

MCP is quickly becoming the way assistants like Alexa+ reach external tools, but the protocol is invisible. When a tool call fails, a response is slow, or a schema is wrong, developers are left guessing or digging through raw logs. We wanted to make that conversation visible.

What it does

MCP Lens has three parts:

  • An MCP server exposing a personal notes workspace through four tools:
    • search_notes: keyword and date-range search
    • summarize: streams its summary back in chunks as it is generated
    • extract_tasks: returns structured tasks (owner, due date, source note) validated against a declared output schema
    • draft_reply: drafts a reply in a friendly, formal or brief tone
  • The Lens, a web view of the live protocol:
    • The handshake (initialize, capability negotiation, initialized, tools/list) as a timeline
    • Every tool call with raw request and response JSON, latency and errors
    • Streaming chunks arriving one by one with timestamps
  • An Alexa+-style chat panel. Ask "What are my open action items from last week's notes?" and the panel calls the tools and shows a task list, while every call lights up in the Lens beside it.

How we built it

  • Transport: MCP over Streamable HTTP on a single endpoint, with bearer-token auth and CORS configured for browser clients.
  • A real client: The web app is a genuine MCP client that runs the full lifecycle and sends real JSON-RPC messages. An instrumented fetch layer captures each request, response and server-sent event, so the Lens shows actual traffic, not a mock-up.
  • Streaming: summarize emits progress notifications while it works and returns the final result at the end. The Lens and chat both render partial output live.
  • Routing: The chat panel maps requests to tool calls. The hook question runs search_notes for last week's date range, then passes the matching note IDs to extract_tasks.
  • Deployment: AWS Lambda behind a Function URL with response streaming, using environment-based configuration. A mock mode with no API keys keeps the demo reliable.
  • Interoperability: Because it follows the spec, other MCP clients (an existing AI assistant, MCP Inspector) can call the same server.
Chat panel ─┐
            ├─> MCP client (instrumented) ──Streamable HTTP──> MCP server (AWS Lambda) ──> notes
Lens view  <┘        every message captured

Real: the MCP server, protocol traffic, streaming, schemas, errors and AWS deployment. Seeded: the notes (sample data). Simulated: the chat panel imitates Alexa+ and does not connect to a physical device.

Challenges we ran into

  • Capturing genuine protocol traffic, including SSE streams, without altering behavior
  • Streaming responses through Lambda Function URLs
  • Ambiguities in the MCP spec and gaps in the SDKs
  • Browser-client concerns: CORS and bearer-token auth
  • Validating structured output against a declared schema

Accomplishments that we're proud of

  • A real, spec-compliant MCP server and client, not a mock-up, deployed on AWS
  • Live visualization of the full handshake, tool calls and streaming chunks
  • Interoperability with other MCP clients such as MCP Inspector
  • A friction log of spec ambiguities, SDK gaps and deployment surprises, offered as feedback to the people building these tools

What we learned

  • MCP is easy to start with but hard to observe; visibility changes how fast you can debug
  • Streaming and progress notifications behave very differently across transports and hosting
  • Declared output schemas catch problems early and make tool results reliable
  • Deployment details (auth, CORS, streaming) matter as much as the protocol itself

What's next for MCP Lens

  • Connect to real Alexa+ integrations and live tool servers beyond the notes demo
  • Let developers point the Lens at any MCP server URL
  • Add request replay, schema diffing and latency analytics
  • Publish the friction log and contribute fixes or issues back to the spec and SDKs

Built With

  • alexa-plus
  • aws-lambda
  • bearer-auth
  • cors
  • json-rpc
  • json-schema
  • lambda-function-url
  • mcp
  • model-context-protocol
  • node.js
  • react
  • server-sent-events
  • streamable-http
  • tailwindcss
  • typescript
  • vite
  • zod
Share this project:

Updates

Submission history