convez.ai
AgenteDevOps & InfraestruturaAvançado
raphaeltm

Simple Agent Manager

raphaeltm/simple-agent-manager

Execute agentes de IA em paralelo na sua própria infraestrutura

Ver no GitHub
BomPluginScore · de 0 a 100

Orquestrador de agentes de IA que executa múltiplos agentes de código em paralelo em contêineres Docker isolados em VMs da sua própria nuvem (Hetzner, Scaleway, GCP). Suporta Claude Code, Gemini, Mistral e outros, com histórico de conversa persistente e controle serverless via Cloudflare Workers.

Instalação

$npx convez-ai raphaeltm/simple-agent-manager

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
73
Forks
12
Acompanhando
1
Saúde
57%
Último commit
há 1 dia
Versão
v2026.11.16
Issues abertas
34
Contribuidores
12

Casos de uso

  • Executar múltiplos agentes de código simultaneamente em ambientes isolados
  • Automatizar tarefas de desenvolvimento em escala (de 5 a 500+ agentes)
  • Manter histórico de conversas e contexto entre workspaces

Veredito

Solução madura e bem arquitetada para quem precisa orquestrar agentes de código em escala com controle total sobre infraestrutura. Custo operacional baixo (~$0.007-0.03/hr por workspace) e suporte a múltiplos modelos de IA, mas exige conhecimento sólido de cloud computing e DevOps para self-hosting.

README

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


What You Get

Agents in parallel, each in a real environment. Every agent runs in its own isolated Docker container on a VM you own — full Linux, Docker, and git, reachable from any browser. Run a handful or hundreds at once, each in a clean workspace.

Bring your own cloud. VMs are provisioned in your own Hetzner, Scaleway, or GCP account and billed directly to you. SAM never stores your cloud provider credentials as platform env vars — they're encrypted per-user. Your agents, your infra, your data.

Bring your own agent. Six harnesses work today: Claude Code, Codex, Gemini, Mistral, OpenCode, and Amp. Use your own API key, your OAuth/subscription token, or the platform proxy.

Chat-first, and it outlives workspaces. Link a GitHub repo, describe a task in natural language, and watch every tool call stream back. Conversations persist at the project level — stop a workspace, spin up a new one weeks later, and your full history is still there.

Serverless control plane. Self-hosted on Cloudflare ($5/mo Workers Paid plan) — the control plane provisions, schedules, and tears down workspaces with nothing to babysit. A workspace costs ~$0.007–0.03/hr compared to $0.18–0.36/hr on GitHub Codespaces.

How It Works

You: "Add rate limiting to the /api/upload endpoint"
         |
    Project Chat (app.{domain})
         |
    Cloudflare Worker API
         |
    TaskRunner -- alarm-driven orchestrator that:
      1. Claims a warm node or provisions a new VM in your cloud
      2. Creates a Docker workspace with your repo
      3. Starts your chosen agent with the task description
      4. Streams agent output back to project chat
         |
    Agent streams results back as it works

Architecture

LayerWhatHow
Control planeAPI, auth, orchestrationCloudflare Workers + D1 + KV + R2
Real-time dataChat messages, activity, sessionsDurable Objects with embedded SQLite (per project)
ComputeWorkspaces running coding agentsVMs in your own cloud (Hetzner, Scaleway, or GCP) with a Go agent managing Docker containers, WebSocket terminal, and auth
Warm poolFast workspace startsCompleted VMs stay warm for 30 min for instant reuse

The control plane is serverless — no servers to manage, no databases to back up. Compute scales to zero when you're not using it.

Repository Structure

apps/
  api/          Cloudflare Worker API (Hono)
  web/          Control plane UI (React + Vite)
  www/          Marketing site, blog & docs (Astro + Starlight)
packages/
  shared/       Shared types and utilities
  providers/    Cloud provider abstraction (Hetzner, Scaleway, GCP)
  cloud-init/   Cloud-init template generator
  vm-agent/     Go VM agent (PTY, WebSocket, MCP tool endpoints)
  ui/           Design system tokens and shared UI components
  terminal/     Shared terminal component

For the full architecture with diagrams, see the Architecture Overview.

Quick Deploy

SAM deploys automatically via GitHub Actions. Fork, configure, push. For the complete setup guide with detailed steps and troubleshooting, see the Self-Hosting Guide.

Prerequisites

Steps

  1. Fork this repository
  2. Create a GitHub Environment named production in your fork's Settings > Environments
  3. Add the required secrets (Cloudflare API token, GitHub App credentials, etc. — see the Self-Hosting Guide for the full list)
  4. Push to main — GitHub Actions provisions all infrastructure, deploys the API + UI, runs migrations, and verifies health

Your instance is live at app.{your-domain}. Users sign in with GitHub and provide their own cloud provider API token (Hetzner, Scaleway, or GCP) to create workspaces.

Development

pnpm install        # Install dependencies
pnpm build          # Build all packages
pnpm test           # Run tests
pnpm typecheck      # Type check
pnpm lint           # Lint
pnpm format         # Format

Build packages in dependency order: shared > providers > cloud-init > api / web.

For local development details, see the Local Development Guide.

Documentation

Full documentation is available at simple-agent-manager.org/docs:

License

AGPL-3.0

Ver o original no GitHub

Bom para

Relacionados