convez.ai
MCP ServerSegurançaIntermediário
undercurrentai

AEGIS Governance MCP

undercurrentai/aegis-mcp

Gateway de governança quantitativa com auditoria para agentes IA

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que implementa governança quantitativa para agentes de IA através de seis portas de decisão (Risco, Lucro, Novidade, Complexidade, Qualidade, Utilidade). Retorna decisões estruturadas (PROCEED/PAUSE/HALT/ESCALATE) com pontuações de confiança, justificativa e trilha de auditoria encadeada por hash, compatível com NIST AI RMF e EU AI Act.

Instalação

$npx convez-ai undercurrentai/aegis-mcp

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
50%
Último commit
há 2 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Avaliar propostas de mudança arquitetural antes de agentes executarem
  • Gerar trilhas de auditoria compliance-ready para decisões de IA
  • Implementar controles de risco em pipelines de engenharia automatizados

Veredito

Ferramenta sólida e pronta para produção que preenche um gap real em governança de agentes IA. O modelo de sandbox local sem signup reduz fricção, enquanto a auditoria encadeada atende requisitos regulatórios genuínos — ideal para equipes que precisam de controles quantificáveis em decisões automatizadas.

README

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

AEGIS Governance — MCP Server

PyPI License: BSL-1.1

Quantitative governance for AI agents and engineering decisions. AEGIS evaluates proposals through six quantitative gates — Risk, Profit, Novelty, Complexity, Quality, Utility — and returns a structured decision (PROCEED / PAUSE / HALT / ESCALATE) with confidence scores, rationale, and a hash-chained audit trail.

Give your agent a decision gate it can call before it acts — and an audit record compliance can actually read (NIST AI RMF, EU AI Act Annex IV).

  • Works immediately, no signup: the local server runs in sandbox mode (10 evaluations/day).
  • 6 local tools (evaluations, risk checks, health, decision history, usage) — 10 on the hosted server.
  • Hosted server with hash-chained audit trails — free Community tier (100 evaluations/month, no credit card).
  • Want to see it before connecting? Try the Advisor in your browser — no install, no signup.

Quickstart (local, no account needed)

pip install "aegis-governance[mcp]"

Claude Code

claude mcp add aegis -- aegis-mcp-server

Cursor (.cursor/mcp.json) / Windsurf / any stdio MCP client:

{
  "mcpServers": {
    "aegis": { "command": "aegis-mcp-server" }
  }
}

VS Code (.vscode/mcp.json):

{
  "servers": {
    "aegis": { "type": "stdio", "command": "aegis-mcp-server" }
  }
}

Runs in sandbox mode out of the box. Set AEGIS_API_KEY in the server's environment (free key) to unlock decision history, usage reports, and risk checks. Requires Python >= 3.10.

Hosted server (streamable-http, full 10-tool surface)

Get a free API key at portal.undercurrentholdings.com (GitHub/Google sign-in, key provisioned automatically), then:

Claude Code

claude mcp add --transport streamable-http aegis https://mcp.aegis.undercurrentholdings.com/mcp \
  --header "Authorization: Bearer YOUR_API_KEY"

Cursor (.cursor/mcp.json) / Windsurf / any streamable-http MCP client:

{
  "mcpServers": {
    "aegis": {
      "type": "streamable-http",
      "url": "https://mcp.aegis.undercurrentholdings.com/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

VS Code (.vscode/mcp.json):

{
  "servers": {
    "aegis": {
      "type": "http",
      "url": "https://mcp.aegis.undercurrentholdings.com/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

Prefer a local SDK instead of MCP?

The Python SDK has a sandbox mode that works with no account at all (10 evaluations/day):

pip install aegis-governance
from aegis import Aegis

decision = Aegis().evaluate(
    proposal_summary="Add Redis caching layer to reduce API latency",
    risk_baseline=0.02, risk_proposed=0.05,
    novelty_score=0.75, complexity_score=0.8, quality_score=0.9,
)
print(decision.status)  # "proceed"

The local stdio MCP server above ships in aegis-governance >= 1.3.0 via the [mcp] extra.

Tools

ToolWhat it does
aegis_evaluate_proposalFull six-gate evaluation of a proposal; returns PROCEED/PAUSE/HALT/ESCALATE with per-gate scores and rationale
aegis_quick_risk_checkFast risk screen for a proposed change
aegis_check_thresholdsCurrent gate threshold configuration
aegis_get_scoring_guideDomain-specific guidance for deriving gate parameters (e.g. cicd)
aegis_record_proposalRecord a proposal for later verification
aegis_list_proposalsList recorded proposals
aegis_verify_proposalsVerify recorded proposals against outcomes
aegis_list_decisionsList past governance decisions
aegis_get_decisionFetch a specific decision with full audit detail
aegis_crypto_statusHash-chain audit integrity status

Why a governance gate?

AI agents make thousands of decisions with no record of why. AEGIS gives every consequential action a quantitative evaluation and a tamper-evident audit entry — so "the agent decided to deploy" becomes a signed, replayable record with gate scores and rationale.

  • Six gates: Risk, Profit, Novelty, Complexity, Quality, Utility — calibrated thresholds, KL-divergence drift detection
  • Audit-ready: hash-chained decision log; NIST AI RMF and EU AI Act Annex IV artifact generation
  • Five integration surfaces: MCP (this repo), Python SDK, REST API, CLI, GitHub Action

Links


Built by Undercurrent — Agency over agents.

Ver o original no GitHub

Bom para

Relacionados