convez.ai
MCP ServerIntegração & APIsAvançado
VKirill

Agent Consult MCP Server

VKirill/agent-consult

Orquestração multi-agente com síntese inteligente via MCP

Ver no GitHub
BomPluginScore · de 0 a 100

Servidor MCP pronto para produção que orquestra múltiplos agentes de IA (Claude, Codex, Anti-Gravity, Mimo, Grok) para consultas colaborativas com síntese por Minimax-M3. Integra-se nativamente com ferramentas MCP como Claude Desktop e Codex CLI, fornecendo respostas consensuadas e otimizadas.

Instalação

$npx convez-ai VKirill/agent-consult

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
1
Forks
0
Acompanhando
0
Saúde
100%
Último commit
há 6 dias
Versão
v1.4.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Consultas técnicas com múltiplas perspectivas de especialistas
  • Síntese de decisões complexas com resolução de conflitos
  • Integração em pipelines de desenvolvimento com isolamento seguro

Veredito

Solução técnica madura para orquestração multi-agente via MCP com foco em consenso e segurança. Excelente para arquitetos de sistemas que precisam de múltiplas perspectivas de IA integradas, mas requer conhecimento profundo de MCP e infraestrutura.

README

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

Agent Consult MCP Server

A production-grade Model Context Protocol (MCP) server for running multi-agent AI consultations (Codex, Claude, Anti-Gravity, Mimo, Grok) with professional synthesis powered by Minimax-M3 on OpenRouter.

License: MIT Node.js MCP TypeScript

💬 Telegram Channel · 🇷🇺 Русский · 🇨🇳 中文 · 🇪🇸 Español · 🇩🇪 Deutsch


📖 Overview & SEO Description

Agent Consult MCP Server is a robust multi-agent orchestration and consensus-building platform built on top of the Model Context Protocol (MCP). It coordinates a panel of virtual experts (Codex for logic and code, Claude for analysis and writing style, Anti-Gravity for creative thinking, Mimo for structural criticism, and Grok for detailed real-time execution analysis) to deliver a unified, highly optimized professional response. The final synthesis is performed by the advanced Minimax-M3 model, ensuring technical accuracy, conflict resolution, and cohesive formatting.

Designed for developers, system architects, and marketing strategists, it brings enterprise-grade AI consensus directly to tools like Claude Desktop, Codex CLI, and other MCP-compatible clients.


🛠️ System Architecture

flowchart TD
    Client[Calling Agent / MCP Client] -->|ask_consultant| MCPServer[Agent Consult MCP Server]
    
    MCPServer -->|Standard Mode: 5 agents| Agents
    MCPServer -->|Security Mode: 1 agent| CodexHigh[Codex \nopenai/gpt-5.5 + High Reasoning]
    
    subgraph Agents [Consultation Agent Pool]
        Codex[Codex \nAlgorithms & Logic]
        Claude[Claude \nAnalysis & Style]
        AGY[Anti-Gravity \nCreative & Out-of-box]
        Mimo[Mimo \nCritique & Structure]
        Grok[Grok \nDetail Search & Action]
    end
    
    Agents -->|Collect Answers| MCPServer
    MCPServer -->|Consolidation| Synthesis[Synthesizer Minimax-M3]
    Synthesis -->|Synthesized Report| MCPServer
    
    CodexHigh -->|Direct Response without Synthesis| MCPServer
    
    MCPServer -->|Final Markdown Report| Client

For a deep dive into the inner workings, refer to the documentation:


✨ Key Features

  1. Sandbox Isolation (Security First)
    • Every local agent runs in its own isolated home directory (~/.agent-consult/homes/) with custom environment variables.
    • Credentials and OAuth tokens are copied securely with 0600 permissions.
    • Prevents recursive tool execution loop and leaks.
  2. Dynamic Role-Based MCP Mapping
    • Agents are equipped with specific tools depending on their active role. Developers get code tools, marketers get search tools, and architects get database tools.
  3. Consensus Synthesis (Minimax-M3)
    • Resolves contradictions between different models.
    • Aggregates ideas and outputs a professional, structured Markdown report.
  4. Liveness Probe & Resilience
    • Multi-agent requests execute in parallel with a configurable timeout.
    • If one agent hangs or crashes, it is gracefully handled while the others finish the job.
    • A liveness probe dynamically extends timeouts for heavy reasoning models if active work is detected.

📋 MCP Tools Reference

The server exposes the following tools:

1. ask_consultant

Run a multi-agent consultation to answer a complex prompt or task.

  • Arguments:
    • question (string, required): Your question or technical task description.
    • role (enum, optional, default: general): Specialist profile. Available: marketer, programmer, system_architect, web_architect, app_architect, security_auditor, qa_engineer, data_engineer, general.
    • custom_role_prompt (string, optional): Overrides the default system prompt for the role.
    • agents (string[], optional): Sub-list of agents to query (e.g., ["codex", "claude"]). Defaults to ["codex", "claude", "agy", "mimo", "grok"].
    • skip_synthesis (boolean, optional, default: false): Skips the consolidation phase and returns raw agent responses.

2. check_agents_status

Verifies connection to OpenRouter, checks current agent status, and returns network latencies.

3. list_available_roles

Returns the list of all configured roles with their descriptions.


⚙️ Configuration (config.json)

The server configuration resides in config.json. You can modify it on the fly:

{
  "openrouter_api_key": "YOUR_OPENROUTER_API_KEY_HERE",
  "timeout_ms": 240000,
  "retry_attempts": 2,
  "agents": {
    "codex": {
      "model": "openai/gpt-5.5",
      "system_prefix": "You are Codex. Your strength is in algorithmic precision and code analysis...",
      "reasoning": {
        "enable": false,
        "reasoning_effort": "medium"
      }
    },
    "grok": {
      "model": "xai/grok-composer-2.5-fast",
      "system_prefix": "You are Grok. Your strength is in search details and practical advice...",
      "reasoning": {
        "enable": false,
        "reasoning_effort": "medium"
      }
    }
  },
  "synthesis": {
    "model": "minimax/minimax-m3",
    "system_prefix": "You are the Synthesis Engine. Synthesize the following expert reports...",
    "reasoning": {
      "enable": false
    }
  }
}

[!TIP] You can also set the API key via the OPENROUTER_API_KEY environment variable. It takes precedence over the config.json value.


📂 Specialist Role Profiles

Role prompts are located in the profiles/ folder. They are read dynamically on each request:


🛡️ Security Auditor Mode & DevSecOps Tools

The server includes a specialized Security Auditor (security_auditor) mode designed for static/dynamic code analysis, vulnerability scanning, and secure architecture modeling:

  • Single High-Reasoning Execution: When the security_auditor role is selected, the server automatically bypasses multi-agent polling and consensus synthesis. It queries a single local Codex agent running the state-of-the-art openai/gpt-5.5 model with maximum reasoning settings (reasoning_effort: "high").
  • Supply Chain & SAST/DAST Integration: The agent is equipped with special DevSecOps tools:
    • sentinel: Orchestrates security tools such as Trivy (for scanning dependencies/container images), Semgrep (Static Application Security Testing - SAST), and OWASP ZAP (Dynamic Application Security Testing - DAST) under a single interface.
    • skylos: Detects hardcoded secrets, API tokens, and vulnerabilities (taint analysis) for JavaScript, TypeScript, Python, and Go codebases.
  • Testing Security Auditor: You can run the integrated DevSecOps audit test using:
    node dist/test-security-auditor.js
    

🚀 Installation & Quick Start

1. Clone & Build

Ensure you have Node.js v20+ and npm installed:

git clone https://github.com/VKirill/agent-consult.git
cd agent-consult
npm install
npm run build

2. Integration with Claude Desktop

Add the server to your Claude Desktop config (located at ~/.config/Claude/claude_desktop_config.json on Linux/macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "agent-consult": {
      "command": "node",
      "args": [
        "/absolute/path/to/agent-consult/dist/index.js"
      ],
      "env": {
        "OPENROUTER_API_KEY": "YOUR_OPENROUTER_API_KEY"
      }
    }
  }
}

3. Integration with Codex CLI (~/.codex/config.toml)

[mcp_servers.agent_consult]
command = "node"
args = ["/absolute/path/to/agent-consult/dist/index.js"]
startup_timeout_sec = 20
env = { OPENROUTER_API_KEY = "YOUR_OPENROUTER_API_KEY" }

👨‍💻 Developer & Author


📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

Ver o original no GitHub

Bom para

Relacionados