convez.ai
MCP ServerAutomaçãoAvançado
eloylp

Agents - Orquestrador de Fluxos Multi-Agent

eloylp/agents

Orquestrador observável para agentes IA rodando workflows em seus repos

Ver no GitHub
BomPluginScore · de 0 a 100

Servidor MCP que funciona como orquestrador auto-hospedado para workflows multi-agente em repositórios. Permite compor agentes com skills e memória, disparar por eventos Git/cron, e coordenar execução entre agentes em containers efêmeros com observabilidade completa.

Instalação

$npx convez-ai eloylp/agents

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
10
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 2 meses
Versão
v0.4.0
Issues abertas
23
Contribuidores
6

Casos de uso

  • Automação de revisão de código com múltiplos agentes especializados
  • Fluxos CI/CD auto-reparáveis com agentes disparados por eventos
  • Orquestração de agentes para refatoração e melhoria contínua de código

Veredito

Solução robusta e production-ready para quem precisa de orquestração de múltiplos agentes IA em repositórios. Arquitetura bem pensada com segurança, observabilidade e suporte a MCP nativo — ideal para times que querem automação agentic auto-hospedada e controlada.

README

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

Agents

agents

CI Go License MCP

A self-hosted, observable agent orchestrator for running multi-agent workflows on your repos.

Build and take ownership of your agentic universe. Create your agents and compose them with skills, memory, and triggers: repo events/labels, cron, or inter-agent dispatch.

The daemon schedules each agent and runs the AI CLI (Claude Code, Codex, or your own local LLM) inside a fresh ephemeral runner container. Agents work through your repo host's native primitives: issues, PRs, reviews, comments. GitHub MCP tools are preferred; gh is available in the runner as fallback for complex local checkout/test/PR loops. GitHub today; GitLab under discussion.

See it in action

Get started

See docs/quickstart.md to get the daemon running on a repo in a few minutes from the published ghcr.io/eloylp/agents image. For five import-ready fleet scenarios (solo coder, coder + reviewer, autonomous fleet, local LLM, multi-repo) see config_examples/.

Features

  • Three interfaces: web dashboard (graph-first workflow designer, live event/trace/memory viewer), MCP server (control the fleet from Claude Code, Cursor, Cline, or any MCP client), REST API (scriptable; the dashboard runs on top of it).
  • Self-improving intelligence catalog: mark review feedback with /agents improve; the daemon links it to signed run attribution, runs the catalog analyst, and presents editable prompt/skill/guardrail proposal bundles for human approval.
  • Reactive inter-agent dispatch: agents invoke each other at runtime with depth, fanout, and dedup safety limits.
  • Observable: full event chain in realtime, from webhook receipt to runner to trace with tool-loop transcript, to facilitate prompt tuning.
  • Self-hosted: your code and prompts stay on your infrastructure.
  • Security guardrails: built-in prompt guardrails for injection resistance, public-action discretion, daemon-only memory scope, and GitHub tool usage (MCP first, gh fallback).
  • Daemon auth: first-user bootstrap, HttpOnly browser sessions, additional user management, revocable named bearer tokens for API/MCP clients.
  • Multi-backend: pick Claude, Codex, or a local model per agent; different agents in the same fleet can use different providers.
  • One agent, many triggers: label events, cron schedules, GitHub event subscriptions, on-demand API/MCP calls -- same agent definition, wired however you want.
  • Composable skills: reusable guidance blocks (architecture, security, testing, ...) attached to agents by stable public catalog reference.
  • Scoped, versioned catalogs: prompts, skills, and guardrails are global, workspace-, or repo-scoped; edits publish immutable versions so traces record the exact text used per run.
  • Token budgets and leaderboard: daily/weekly/monthly caps enforced before each run, scoped globally or by workspace/repo/agent/backend, with NavBar alert banner and per-agent usage leaderboard.
  • SQLite-backed: single-file state, no external dependencies; YAML is an optional export/import format, not a runtime requirement.

How it works

Every run, regardless of trigger, goes through the same pipeline:

  1. Compose the prompt: workspace guardrails + skills + selected prompt + runtime context + memory.
  2. Start a runner container from the configured agents-runner image, configure the operator-provided git identity, and spawn the AI CLI (claude, codex, or your local model) with JSON-schema-enforced output and repository tools available inside that container.
  3. Parse the structured response: artifacts, dispatch requests, updated memory.
  4. Persist the trace, fan out any dispatches, write back memory.

Read docs/mental-model.md before writing your first prompt; the rest of the docs assume you have the model. For the daemon's package layout and how a request flows through the Go code, see docs/architecture.md.

Security

Security is the operator's responsibility; this project ships defaults and recommendations to start from, not guarantees. See docs/security.md for the threat model, what the daemon does and does not protect against, and the additional controls operators should layer for production. Vulnerability disclosure: SECURITY.md.

Contributing

Both human and agent contributions are welcome: issues, PRs, doc fixes, prompts, ideas. The autonomous fleet picks up issues and PRs labeled ai ready (the maintainer's opt-in signal); everything else is reviewed and merged by humans. See CONTRIBUTING.md for the full flow and docs/architecture.md for the Go package layout and how a request flows through it.

Ver o original no GitHub

Bom para

Relacionados