Skip to main content
Glama
jherr

Fretboard Voicing Explorer MCP Server

by jherr

Fretboard Voicing Explorer — MCP Apps dual-host demo

A runnable demo of MCP Apps portability: a TanStack Start app hosts its own HTTP MCP server, consumes that server to render an interactive fretboard ui:// widget inline in chat, and the same server is consumed by Goose to render the same widget in a different host.

One MCP server. One ui:// resource. Two hosts.

What's here

src/
  mcp/
    voicing.ts                # pure music model: chord→voicing, voice-leading analysis
    voicing.test.ts           # vitest unit tests (deterministic)
    server.ts                 # MCP server: show_voicing, evaluate_voicing, analyze_progression
    widget/fretboard.html     # host-agnostic ui:// widget (self-contained, no app imports)
  routes/
    fretboard.tsx             # chat host UI + <MCPAppResource> render
    api.mcp.ts                # the Streamable HTTP MCP server endpoint
    api.fretboard.chat.ts     # chat route: createMCPClient(self) + Anthropic adapter
    api.mcp-app.call-tool.ts  # bridge call endpoint (allowlisted)
  sandbox/sandbox_proxy.html  # separate-origin MCP Apps sandbox proxy (served on :3100)
scripts/sandbox-server.mjs    # tiny static server for the sandbox proxy
goose/config-snippet.md       # how to point Goose at /api/mcp

Related MCP server: guitar-pro-mcp

Prerequisites

  • Node 20+, pnpm

  • ANTHROPIC_API_KEY in .env.local (the chat route's Anthropic adapter reads it)

  • Goose Desktop ≥ v1.19.0 for the cross-host half (verified on v1.39.0)

Run

pnpm install
cp .env.example .env            # ports: app 4321, sandbox proxy 3100
echo "ANTHROPIC_API_KEY=sk-ant-..." > .env.local

pnpm dev                        # starts the app (:4321) AND the sandbox proxy (:3100)

Open http://localhost:4321/fretboard. The app uses a fixed port (4321, set in vite.config.ts with strictPort) so the origin is stable for Goose; if it's taken, free it or change the port in vite.config.ts + Goose config.

pnpm test                       # runs the pure-model unit tests

Architecture

  Browser (host #1) ──▶ /fretboard          chat UI, renders <MCPAppResource>
                        /api/fretboard/chat  createMCPClient(self) + Anthropic
   ┌── HTTP MCP ──────▶ /api/mcp             MCP server (3 tools + ui:// resource)
   │                    /api/mcp-app/call-tool  bridge endpoint (evaluate_voicing only)
   │
   │   widget iframe ──▶ :3100 sandbox proxy  separate origin (security isolation)
   │
   └── Goose (host #2) ── points at /api/mcp ── renders the SAME ui:// widget

The ui:// widget speaks the mcp-ui postMessage protocol; the server wraps it with the mcp-apps adapter (@mcp-ui/server) which translates to the MCP Apps (ext-apps) protocol both hosts speak — so the widget is host-agnostic.

Resolve-first findings (verified from source)

  1. ui:// reconciliation is a pull model. @tanstack/ai-mcp@0.2.0 reads the tool's _meta.ui.resourceUri at discovery, stamps it, then fetches the resource via resources/read and matches content.uri === uiResourceUri. So show_voicing sets _meta.ui.resourceUri (and openai/outputTemplate for Goose/Apps SDK) and the server exposes the resource via resources/read.

  2. Adapter: anthropicText('claude-sonnet-4-6') from @tanstack/ai-anthropic.

  3. Transport: the MCP SDK's WebStandardStreamableHTTPServerTransport (RequestResponse, stateless per-request) serves /api/mcp straight from a TanStack Start route — no Node bridging.

  4. @mcp-ui/server@6.1.0 wrapHtmlWithAdapters + getAdapterMimeType emit text/html;profile=mcp-app.

  5. Sandbox proxy is the ext-apps double-iframe proxy, self-hosted on a separate origin (:3100) — @mcp-ui/client@7.1.1's AppRenderer requires it.

Demo script

  1. TanStack app. /fretboard → "Show me a Cmaj7 voicing." The fretboard renders inline. Drag a note → evaluate_voicing → the analysis updates. "What comes next?" → sends a prompt that advances the progression.

  2. Goose, same server. Follow goose/config-snippet.md, restart Goose, ask for the same chord. The same widget renders in Goose; the same drag yields the same evaluate_voicing result; narration comes from Goose's own model.

  3. Honest divergence + security. Same tool result, host-specific narration. A javascript: link is rejected by both hosts' scheme allowlist. A text-only client still gets a usable voicing via the mandatory text fallback.

Status

  • ✅ One /api/mcp server with 3 tools + the ui:// resource (verified over the wire).

  • ✅ TanStack host renders the ui:// widget cross-origin through the sandbox proxy.

  • show_voicing runs, the model narrates the computed voicing; text fallback present.

  • ✅ Pure evaluate_voicing model is unit-tested and deterministic.

  • ✅ Live in-widget host bridge works: the widget is seeded with the computed voicing (delivered as ui/notifications/tool-input), and an in-widget tool call round-trips evaluate_voicing through the bridge endpoint and renders the result.

  • ✅ Bridge endpoint enforces an allowlist (evaluate_voicing only).

Isolated widget harness

/fretboard-test renders the widget with a hardcoded Cmaj7 seed and the real bridge — no chat, no API key — handy for working on the widget in isolation.

Notes / gotchas

  • The widget HTML is inlined into the server module via Vite ?raw, so after editing src/mcp/widget/fretboard.html you need a full dev restart (HMR alone serves a stale widget).

  • The bridge result comes back double-wrapped (the bridge returns a full CallToolResult and the host wraps it again); the widget unwraps to the analysis. See structuredFrom in the widget.


This app was scaffolded with TanStack Start (file-based routing, Nitro server, Tailwind). See https://tanstack.com/start for framework docs.

F
license - not found
-
quality - not tested
B
maintenance

Maintenance

Maintainers
Response time
Release cycle
Releases (12mo)
Commit activity

Resources

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/jherr/ts-ai-mcp-apps-demo'

If you have feedback or need assistance with the MCP directory API, please join our Discord server