convez.ai
MCP ServerIntegração & APIsAvançado
swannekim

MCP Curator Servers - Agente de Curadoria de Museu

swannekim/mcp-curator-servers

Servidor MCP para curadoria de arte e imagens espaciais em Copilot Studio

Ver no GitHub
InicialPluginScore · de 0 a 100

Conjunto de servidores MCP customizados para o Copilot Studio que implementa um agente de curadoria de arte e espaço. Integra APIs da NASA e do Art Institute of Chicago para criar exibições temáticas e encontrar correspondências artísticas baseadas em personalidade, com geração de visualizações Mermaid.

Instalação

$npx convez-ai swannekim/mcp-curator-servers

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Criar exibições temáticas combinando arte com imagens espaciais
  • Encontrar correspondências de obras de arte baseadas em palavras-chave de personalidade
  • Gerar visualizações e ranking de obras com Mermaid
  • Integrar dados da NASA e Art Institute em agentes de IA

Veredito

Projeto maduro e bem documentado que demonstra padrões sólidos de construção de MCP servers em TypeScript para Copilot Studio. Excelente referência para arquitetura multi-servidor com integração de APIs externas e geração de conteúdo visual, mas voltado especificamente ao ecossistema Microsoft.

README

A tradução para o português deste README ainda está na fila.

Copilot Studio Curator MCP Servers

This repo contains the source code and setup docs for the MCP servers used by a new Copilot Studio agent.

The agent is built for the New Copilot Studio experience announced on June 2026.

That matters because this project does not use classic Copilot Studio topics. The agent uses:

  • MCP tools for external actions and data.
  • Markdown skills stored as SKILL.md files in docs/skills.
  • A Copilot Studio export in docs/Museum-Curator-Agent.yaml, where the skills appear as InlineAgentSkill components and the MCP servers appear as McpTool components.

The custom MCP servers are deployed to Azure Container Apps by scripts/deploy-azure-containerapps.ps1.


What This Agent Does

The exported agent has two reusable skills:

  • Museum of the Universe Curator - builds a 3-5 piece mini exhibition that pairs NASA space imagery with public-domain Art Institute of Chicago artworks.
  • AI Curator: Art Doppelganger - makes a playful metadata-based artwork match from 3-5 personality words and optional preferred colours.

Both skills can ask the custom scoring server to generate Mermaid syntax, then ask the hosted Mermaid Chart MCP server to render it.


Repo Layout

mcp-curator-servers/
  docs/
    Museum-Curator-Agent.yaml       # Copilot Studio agent export
    copilot-studio-setup.md         # Step-by-step setup notes
    instructions/                   # Older/pasteable instruction drafts
    skills/
      museum-of-universe/
        SKILL.md                    # New Copilot Studio skill markdown
      art-doppelganger/
        SKILL.md                    # New Copilot Studio skill markdown
  scripts/
    deploy-azure-containerapps.ps1  # Builds and deploys all custom MCP servers
  servers/
    nasa-mcp/                       # NASA APOD + Image Library MCP server
    artic-mcp/                      # Art Institute of Chicago MCP server
    curator-scoring-mcp/            # Pairing/ranking + Mermaid syntax server
  package.json                      # Convenience scripts for all servers
  README.md

MCP Servers

There are three custom Streamable HTTP MCP servers in this repo.

ServerLocal portMain tools
nasa-mcp3001get_space_images_for_theme, get_apod_image, get_space_theme_keywords
artic-mcp3002find_artworks_by_mood, get_artwork_detail
curator-scoring-mcp3003pair_space_images_with_artworks, rank_artworks_for_personality, build_exhibition_mermaid, build_doppelganger_mermaid

Mermaid rendering is not implemented in this repo. Copilot Studio connects directly to the hosted Mermaid Chart MCP server:

https://mcp.mermaid.ai/mcp

The scoring server only generates Mermaid syntax.


Agent Definition

The full Copilot Studio agent export is here:

docs/Museum-Curator-Agent.yaml

Inside that YAML:

  • NASA Space MCP, Art Institute CHI MCP, Scoring MCP, and Mermaid Chart MCP are McpTool components.
  • museum-of-universe and art-doppelganger are InlineAgentSkill components.
  • The inline skill content comes from the markdown skills under docs/skills.

The source skill files are:

docs/skills/museum-of-universe/SKILL.md
docs/skills/art-doppelganger/SKILL.md

Update those SKILL.md files first when you want to change the agent behavior. Then re-import or update the skills in Copilot Studio so the YAML/export matches the repo.


How The Tool Chain Works

Copilot Studio is the orchestrator. The MCP servers do not call each other.

Museum skill:

User theme
  -> NASA Space Data MCP: get_space_images_for_theme
  -> Art Institute Public Artwork MCP: find_artworks_by_mood
  -> Curator Scoring MCP: pair_space_images_with_artworks
  -> Curator Scoring MCP: build_exhibition_mermaid
  -> Mermaid Chart Rendering MCP: render/validate Mermaid syntax

Art Doppelganger skill:

User personality words + optional colours
  -> Art Institute Public Artwork MCP: find_artworks_by_mood
  -> Curator Scoring MCP: rank_artworks_for_personality
  -> Curator Scoring MCP: build_doppelganger_mermaid
  -> Mermaid Chart Rendering MCP: render/validate Mermaid syntax

The outputs are creative curation results, not scientific or psychological claims.


Prerequisites

  • Node.js 22+
  • npm
  • Azure CLI, for Azure Container Apps deployment
  • A NASA API key from https://api.nasa.gov
  • An Art Institute of Chicago user agent string, such as museum-universe-curator/you@example.com

Run Locally

Install dependencies for all three servers:

npm run install:all

Create local env files:

Copy-Item servers\nasa-mcp\.env.example servers\nasa-mcp\.env
Copy-Item servers\artic-mcp\.env.example servers\artic-mcp\.env
Copy-Item servers\curator-scoring-mcp\.env.example servers\curator-scoring-mcp\.env

Set at least:

servers/nasa-mcp/.env
  NASA_API_KEY=your_real_key_here
  PORT=3001

servers/artic-mcp/.env
  AIC_USER_AGENT=museum-universe-curator/you@example.com
  PORT=3002

servers/curator-scoring-mcp/.env
  PORT=3003

Run each server in a separate terminal:

npm run dev:nasa
npm run dev:artic
npm run dev:scoring

Health checks:

curl http://localhost:3001/healthz
curl http://localhost:3002/healthz
curl http://localhost:3003/healthz

Each MCP endpoint is:

http://localhost:3001/mcp
http://localhost:3002/mcp
http://localhost:3003/mcp

Test The MCP Servers

The easiest local test tool is MCP Inspector:

npx @modelcontextprotocol/inspector

Use:

  • Transport: Streamable HTTP
  • NASA URL: http://localhost:3001/mcp
  • Art URL: http://localhost:3002/mcp
  • Scoring URL: http://localhost:3003/mcp

Quick raw JSON-RPC smoke test:

curl -X POST http://localhost:3001/mcp `
  -H "Content-Type: application/json" `
  -H "Accept: application/json, text/event-stream" `
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Deploy To Azure Container Apps

The deployment script builds all three server images with Azure Container Registry and deploys them as externally reachable Azure Container Apps.

$env:NASA_API_KEY = "your_real_key_here"
$env:AIC_USER_AGENT = "museum-universe-curator/you@example.com"

.\scripts\deploy-azure-containerapps.ps1

By default, the script creates or updates:

  • Resource group: rg-mcp-curator
  • Container Apps environment: cae-mcp-curator
  • Azure Container Registry: generated from mcp-curator + suffix
  • Container apps:
    • mcp-curator-nasa-sage
    • mcp-curator-artic-sage
    • mcp-curator-scoring-sage

The script stores NASA_API_KEY and AIC_USER_AGENT as Container Apps secrets and prints:

Health checks:
  https://<nasa-app>/healthz
  https://<artic-app>/healthz
  https://<scoring-app>/healthz

Copilot Studio MCP URLs:
  NASA:    https://<nasa-app>/mcp
  ArtIC:   https://<artic-app>/mcp
  Scoring: https://<scoring-app>/mcp

Useful overrides:

.\scripts\deploy-azure-containerapps.ps1 `
  -ResourceGroup "rg-my-curator" `
  -Location "eastus" `
  -NamePrefix "mcp-curator" `
  -UniqueSuffix "demo"

Register Tools In Copilot Studio

In the new Copilot Studio agent designer, add these as MCP tools. Use Streamable HTTP transport.

Tool name in Copilot StudioURLAuth
NASA Space Data MCPhttps://<nasa-app>/mcpNone
Art Institute Public Artwork MCP or Art Institute CHI MCPhttps://<artic-app>/mcpNone
Curator Scoring MCP or Scoring MCPhttps://<scoring-app>/mcpNone
Mermaid Chart Rendering MCP or Mermaid Chart MCPhttps://mcp.mermaid.ai/mcpNone

Good server descriptions:

  • NASA: Use only to retrieve NASA space imagery, astronomy descriptions, and space keywords. Not for artwork or diagrams.
  • Art Institute: Use only to search public-domain artworks and retrieve artwork metadata and IIIF image URLs.
  • Curator Scoring: Use only to pair space images with artworks, rank artworks for a personality, and generate Mermaid diagram syntax. Does no external lookups.
  • Mermaid Chart: Use only to validate and render complete Mermaid diagram syntax into a PNG/SVG/playground link.

Add Skills In Copilot Studio

This repo uses skills, not topics.

Add or update these markdown skills in the new Copilot Studio agent designer:

docs/skills/museum-of-universe/SKILL.md
docs/skills/art-doppelganger/SKILL.md

The new Copilot Studio experience exposes instructions, skills, tools, and knowledge together in the agent builder. These skill files are reusable instructions that load when the user's request matches the skill description.

The older files in docs/instructions are useful reference drafts, but the canonical behavior for this repo is in docs/skills/*/SKILL.md and the exported agent YAML.


Demo Prompts

Museum skill:

Curate a 5-piece exhibition about loneliness in space.
Make a universe museum about wonder.
Build an exhibition on chaos and cosmic mystery.

Art Doppelganger skill:

Find my art doppelganger. I am analytical, romantic, introverted, and secretly dramatic. I like blue and silver.
I am calm, curious, and nostalgic. Match me with an artwork. I like green and gold.
What artwork am I? Ambitious, chaotic, dramatic.

Troubleshooting

SymptomFix
Copilot Studio cannot connectConfirm the Container App URL ends in /mcp, not /healthz, and verify https://<app>/healthz works first.
Tool is registered but not calledCheck that generative orchestration is enabled and the MCP tool description is narrow.
Agent behaves like the old topic versionUpdate the Copilot Studio skills from docs/skills/*/SKILL.md; do not paste the old topic-flow section from an older README.
Scoring tool fails on inputKeep result sets small. The Museum skill defaults to about 5 NASA images and 6 artworks.
Mermaid render failsReturn the curator result anyway, then shorten labels or curator statements before retrying the render.
NASA returns rate limitsUse a real NASA_API_KEY, not DEMO_KEY.
Art search returns weak resultsBroaden the mood term once and prefer public-domain artworks with image URLs.

Suggested Build Order

  1. Run the three MCP servers locally and test them with MCP Inspector.
  2. Deploy them with scripts/deploy-azure-containerapps.ps1.
  3. Register the three deployed custom MCP URLs plus the hosted Mermaid Chart MCP URL in Copilot Studio.
  4. Add the two skills from docs/skills.
  5. Import or compare against docs/Museum-Curator-Agent.yaml.
  6. Test the Art Doppelganger skill first because it has the shorter chain.
  7. Test the Museum skill once NASA, Art, Scoring, and Mermaid all connect.
Ver o original no GitHub

Bom para

Relacionados