convez.ai
MCP ServerIntegração & APIsAvançado
herrbasan

MCP Server Orchestrator

herrbasan/mcp_server

Orquestrador MCP multi-agente com automação web, GitHub relay e memória semântica

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP centralizado que hospeda múltiplos agentes especializados (browser, GitHub, memória, vision) e expõe uma interface unificada via Server-Sent Events. Integra automação web com Puppeteer, análise de código, pesquisa semântica e consolidação autônoma de memória em intervalos regulares.

Instalação

$npx convez-ai herrbasan/mcp_server

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
14%
Último commit
há 5 dias
Versão
—
Issues abertas
5
Contribuidores
1

Casos de uso

  • Automação de pesquisa web com navegador persistente e análise de visão
  • Integração com repositórios GitHub para navegação de código, PRs e issues sem checkout local
  • Consolidação autônoma de memória com embeddings vetoriais entre sessões

Veredito

Projeto robusto de orquestração MCP com arquitetura modular bem pensada. Requer setup externo (LLM Gateway em localhost:3400) e oferece capacidades avançadas de automação e memória persistente, ideal para fluxos de pesquisa e análise complexos.

README

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

MCP Server Orchestrator

A centralized Model Context Protocol (MCP) server that hosts multiple domain-specific agents and exposes a single unified interface to clients via Server-Sent Events (SSE).

Features

  • Agent architecture — domain-specific modules (browser, docs, github, inspector, llm, memory, research, vision) that own their own state and tools
  • LLM Gateway integration — thin WebSocket client to an external LLM Gateway at localhost:3400
  • Browser automation — persistent Puppeteer browser with session management for web research and scraping
  • Semantic memory — vector embedding-based recall across sessions
  • Dreaming system — hourly autonomous memory consolidation into a structured Map with clusters, bridges, and progressive compression
  • GitHub relay — browse code, history, PRs, issues from any GitHub repo without local checkout
  • Vision analysis — iterative image analysis with drill-down focus capability

Setup

1. Install dependencies

npm install

2. Configure environment

Copy .env.example to .env and set the gateway endpoints:

GATEWAY_URL=ws://localhost:3400/v1/realtime
GATEWAY_HTTP_URL=http://localhost:3400

3. Configure environment (continued)

Model routing is handled by the LLM Gateway. See the Gateway's configuration for task-to-model mapping.

4. Start the server

npm start

The MCP SSE endpoint is available at http://<HOST>:<PORT>/sse.

5. Connect a client

In VS Code mcp.json:

{
    "servers": {
        "orchestrator": {
            "type": "sse",
            "url": "http://localhost:3100/sse"
        }
    }
}

Architecture

Client (VS Code / CLI)
        | SSE (custom transport)
        v
   MCP Orchestrator  (src/server.js)
        | WebSocket
        v
   LLM Gateway  (external — localhost:3400)

Key files:

FilePurpose
src/server.jsExpress + custom SSE transport, JSON-RPC routing
src/agent-loader.jsScans src/agents/, auto-loads prompts, init/shutdown lifecycle
src/gateway-client.jsWebSocket + HTTP client to the LLM Gateway
src/agents/One folder per agent domain

Configuration:

FileContents
.envSecrets and runtime endpoints (GATEWAY_URL, GATEWAY_HTTP_URL, GIT_TOKEN, MAX_MEMORY_CHARS)
config.jsonNon-secret settings: agent options, port binding

Agents

AgentToolsDescription
browserbrowser_session_create, browser_session_goto, browser_session_click, … (14 tools)Persistent Puppeteer browser with session management
docsget_philosophy, get_orchestrator_docAccess mcp_documentation/ files
dreamingdream_generate, dream_status, dream_injectAutonomous memory consolidation into structured Map
githubgit_read_file, git_list_tree, git_log, git_get_commit, git_diff, git_list_branches, git_search_repos, git_search_code, git_search_issues, git_repo_info, git_pr_list, git_get_pr, git_issue_list, git_get_issue, git_create_issue (15 tools)GitHub repo relay — browse code, history, PRs, issues from any repo
inspectorinspect_codeLLM-based code review with file loading
llmquery_modelDirect LLM queries via Gateway
memorymemory_store, memory_recall, memory_get, memory_forget, memory_list, memory_update, memory_overviewSemantic vector memory
researchresearch_topicMulti-phase web research pipeline
visionvision_create_session, vision_analyze, vision_list_sessions, vision_get_session, vision_close_sessionIterative image analysis with drill-down

License

MIT

Ver o original no GitHub

Bom para

Relacionados