convez.ai
MCP ServerSegurançaAvançado
mizcausevic-dev

MCP Tool Cards

mizcausevic-dev/mcp-tool-card-spec

Documentos de divulgação estruturados para ferramentas MCP

Ver no GitHub
InicialPluginScore · de 0 a 100

Especificação para documentos de divulgação por ferramenta em servidores MCP, informando efeitos colaterais, compatibilidade com LLMs, latência, custo e superfície de auditoria. Camada sobre MCP que não modifica o protocolo, preenchendo gaps de segurança e operacional.

Instalação

$npx convez-ai mizcausevic-dev/mcp-tool-card-spec

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
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Avaliação de segurança e compliance antes de adotar um MCP server
  • Comunicação clara de efeitos colaterais e riscos entre equipes
  • Matriz testada de compatibilidade com modelos de IA

Veredito

Especificação bem pensada que resolve um problema real: MCP não padroniza divulgação de efeitos colaterais, perfil de risco ou matriz de testes. Tool Cards é uma camada compatível e não-invasiva, essencial para revisores de segurança e engenheiros de plataforma aprovarem adoção de servidores MCP em produção.

README

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

MCP Tool Cards

Validate examples License: MIT

A draft specification for per-tool disclosure documents describing what an MCP tool does, what it can break, what it has been tested with, and how to report misbehavior.

The Model Context Protocol gives clients a way to discover an MCP server's tool surface. It does not give the server a way to disclose each tool's side effects, safety posture, latency profile, or tested-LLM compatibility. MCP Tool Cards fill that gap.

One JSON document per tool. Discoverable. Machine-readable. Designed for the audience that has to approve an MCP server's adoption — security reviewers, platform engineers, and CISOs.

The three pillars

PillarWhat it does
SchemaThe input/output contract — either inline JSON Schema or a pointer to the MCP-native schema served by the tool
SafetySide-effect classification, refusal modes, PII exposure profile, secret-handling profile, human-approval requirements
TrialTested-LLM matrix with pass rates, performance characteristics (p50/p99 latency, cost per call), and an audit-log surface

Why doesn't MCP already cover this?

The MCP specification standardizes discovery and invocation — how a client lists available tools and calls them. It does not standardize:

  • Side effects classification (a read_database tool and a delete_user tool both look identical in MCP)
  • Tested-LLM matrix — which models have been validated, with what pass rates
  • Latency and cost expectations — operators need this to plan capacity and budget
  • PII / secrets exposure — security reviewers need a structured answer, not prose
  • Refusal behavior — when should the tool refuse to execute, and what does refusal look like

Tool Cards layer on top of MCP without modifying it. An MCP server that publishes Tool Cards remains MCP-compliant; clients that don't read Tool Cards continue to work.

Quickstart

  1. For each tool exposed by your MCP server, author a Tool Card conforming to tool-card.schema.json.
  2. Validate with any JSON Schema 2020-12 validator.
  3. Serve at https://<your-mcp-server>/.well-known/mcp-tools/<tool-name>.json and reference the Tool Card URI from your MCP server's tool list.
  4. Pair with an Agent Card on consuming agents — agents reference Tool Cards via capabilities.tools[].mcp_tool_card_uri.

Files in this repo

  • SPEC.md — full v0.1 specification
  • tool-card.schema.json — JSON Schema (draft 2020-12)
  • examples/ — reference Tool Cards for a read-only billing lookup, a mutating ticket-create, and a destructive admin reset

Status

v0.1 draft. Issues and pull requests welcome.

License

MIT-licensed. The specification text, JSON Schema, and example documents in this repository may be freely implemented, extended, redistributed, or incorporated into commercial or non-commercial products with attribution. Reference implementations of this spec (such as mcp-kinetic-gain) are licensed separately under AGPL-3.0.

Kinetic Gain Protocol Suite

A family of open specifications for the answer-engine era. Each spec is a self-contained JSON document format with its own JSON Schema and reference examples; together they compose into an end-to-end account of entity, agent, prompt, tool, and citation.

SpecWhat it does
AEO ProtocolEntity declaration at /.well-known/aeo.json — authoritative claims, citation preferences, audit hooks
Prompt ProvenanceVersioned, lineaged, reviewable LLM prompt records
Agent CardsDeclarative agent capability and refusal disclosure
AI Evidence FormatStructured citations that travel with LLM-generated claims
MCP Tool CardsPer-tool disclosure layered on Model Context Protocol servers

Connect: LinkedIn · Kinetic Gain · Medium · Skills

Ver o original no GitHub

Bom para

Relacionados