convez.ai
MCP ServerAutomaçãoIntermediário
chapmanjw

Rutherford — Plugin de Orquestração de Agentes

chapmanjw/rutherford-claude-plugin

Dê a seu agente IA uma equipe de codificadores coordenados

Ver no GitHub
BomPluginScore · de 0 a 100

Plugin para Claude Code que auto-registra o servidor MCP Rutherford, permitindo orquestrar múltiplos agentes de codificação (Claude Code, Codex, Cursor, Goose) via Agent Client Protocol. Oferece skills, um agente orquestrador e comandos slash para delegar tarefas, executar consenso, debate técnico e revisão de código em paralelo ou sequencial.

Instalação

$npx convez-ai chapmanjw/rutherford-claude-plugin

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
1
Acompanhando
0
Saúde
71%
Último commit
há 5 dias
Versão
v0.2.4
Issues abertas
1
Contribuidores
1

Casos de uso

  • Orquestração de múltiplos agentes de IA para tarefas de codificação complexas
  • Revisão de código multi-modelo com debate e consenso entre agentes
  • Delegação inteligente de subtarefas a diferentes agentes especializados
  • Coordenação paralela de agentes para acelerar desenvolvimento

Veredito

Solução robusta para arquitetar fluxos de trabalho com múltiplos agentes IA em ambientes de desenvolvimento. Excelente para quem precisa coordenar agentes heterogêneos, mas requer familiaridade com MCP e Agent Client Protocol para máximo proveito.

README

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

Prerequisite: uv on your PATH (install it).

/plugin marketplace add chapmanjw/rutherford-claude-plugin
/plugin install rutherford@rutherford-claude

Then restart Claude Code. The skills appear under /rutherford:*, the rutherford-orchestrator agent becomes available, and the Rutherford MCP server is registered automatically.

After installing, you can run a whole Claude session as the orchestrator. Start Claude with the agent flag and no prompt:

claude --agent rutherford:rutherford-orchestrator

Sam opens on the crew menu automatically, health-checks the crew with doctor, and routes each request to the right mode (delegate, consensus, debate, review, or plan). Pass a task on the command line instead and he routes it straight away. In an ordinary Claude Code session the same agent is also available as a subagent for a "just use Rutherford for this" ask.

What this plugin is

Rutherford gives your coding agent a crew. This is the Claude Code-facing piece: it auto-registers the Rutherford MCP server — a stdio server that speaks the Agent Client Protocol to each coding agent — and adds the skills, an orchestrator agent, and slash commands that drive it.

From inside Claude Code you hand one task to one agent, ask several the same question in parallel, have them argue across rounds, or run a multi-model code review, using agents you already log into. Rutherford reuses each agent's own login and never calls a model provider's API.

The plugin launches the server with uvx rutherford-mcp-server, which fetches it from PyPI on first run and caches it, so there is no separate install step. If you would rather install the server yourself, uv tool install rutherford-mcp-server (or pipx install rutherford-mcp-server) puts a rutherford-mcp-server command on your PATH; the setup-rutherford skill covers that path.

See it work

Ask Codex, Claude Code, and Gemini where the deadlock in queue.py is, and show me all three.

One consensus call fans the question out to three agents, each in its own ACP session, and returns every voice. Run as Rutherford, in character:

.---------.
|  \/\/\/ |    Sam Rutherford here — oh, a deadlock hunt, this is a fun one.
|  O  [==]|    Putting the crew on queue.py now.
|    <    |
|  \___/  |
'---------'

consensus   claude_code, codex, gemini
  claude_code   enqueue() holds the lock across the await on flush() — flush() re-enters it
  codex         same root cause; take the lock off the await, or give flush() its own lock
  gemini        classic producer/consumer wait on the shared mutex; reorder the acquire

The persona is flavor on top of accurate work, never instead of it: a failed health check, a refused write, or a real warning is always reported straight. Voice and guardrails live in reference/persona.md.

Quickstart

/rutherford:setup                                   # verify the connection, run doctor, scaffold config
/rutherford:doctor                                  # which agents actually spawn, handshake, and answer
/rutherford:consensus where is the deadlock in queue.py?

Bring your own crew. Rutherford never installs or logs in to an agent; it drives the ones you already run. You need at least two ACP-capable agents installed and signed in (Claude Code and Codex are enough), or a local model: with Ollama or LM Studio running, Rutherford auto-detects each tool-capable model and registers it as a free voice.

What's inside

Skills

Setup and configuration:

SkillWhat it does
setup-rutherfordVerify the connection, run doctor, scaffold config.toml, install missing ACP adapters.
configure-panelsDefine reusable crews (panels) in panels.toon, then hot-reload them.
configure-defaultsSet config.toml defaults: safety posture, trusted workspaces, effort, persistence.
configure-rolesAuthor custom role personas, with the server-restart caveat.
add-agentsDiscover installed ACP agents, adopt a local model, or define an agent by hand.
configure-permissionsAllowlist the Rutherford tools and .rutherford directories so they stop prompting.
troubleshoot-connectionDiagnose why an agent will not drive, from doctor states to fixes.

Driving the crew:

SkillWhat it does
delegate-taskHand one task to one agent (read-only by default).
multi-agent-consensusAsk several agents the same question in parallel; compare or vote.
agent-debateA multi-round debate where agents see each other's positions and revise.
code-review-panelReview a diff or changed files across agents at a senior/principal bar.
background-jobsRun long work as a background job; manage and persist it.
safe-write-delegationLet an agent edit code, behind the trusted-workspace gate and a sandbox.

Agent

rutherford-orchestrator — health-checks the crew with doctor, then routes a request to the right mode: delegate, consensus, debate, review, or plan. Use it for a "just use Rutherford for this" ask, or for a fresh, independent multi-agent take on your own work. Run an entire session as this agent with claude --agent rutherford:rutherford-orchestrator, or invoke it as a subagent from any session.

Slash commands

/rutherford:setup, /rutherford:doctor, /rutherford:panels, /rutherford:permissions, /rutherford:consensus, /rutherford:debate, /rutherford:review.

Examples

examples/panels.toon (a code-review, design-roundtable, and ship-vote panel), examples/roles/principal-reviewer.md, and examples/config.toml — copy one and edit.

The MCP server

This plugin is the Claude-facing piece. The orchestration lives in rutherford-mcp-server (PyPI), a stdio MCP server that speaks ACP to each coding agent. It exposes 18 tools and 19 built-in agents.

Claude Code
   |  MCP over stdio
rutherford-mcp-server          (the ACP client)
   |  ACP over stdio, one session per voice
   +--> claude-agent-acp, codex-acp, goose acp, cursor-agent acp, ... 19 built-in agents

Safety

Everything defaults to read-only. The write and yolo modes are explicit opt-in behind a trusted-workspace gate, and the deliberation tools (consensus, debate, review, plan) never write. See reference/safety.md and the safe-write-delegation skill.

Permissions

A plugin cannot grant Claude Code permissions, so out of the box Claude prompts before editing files in the .rutherford directories and before each Rutherford tool call. The configure-permissions skill (also run as a step of /rutherford:setup) writes a recommended allowlist into your settings.json with your confirmation: read/write access to the project and home .rutherford directories, plus the safe config and inspection tools. It leaves the agent-spawning tools prompting unless you opt in, and it does not weaken Rutherford's own write gate. See reference/permissions.md.

Documentation

The plugin bundles reference docs that the skills cite: tools, safety, panels, config, permissions, and persona. For the full server reference, see the rutherford-mcp-server docs.

The name

.---------.
|  \/\/\/ |
|  O  [==]|
|    <    |
|  \___/  |
'---------'
-- Ensign Sam Rutherford --
USS Cerritos . Engineering

Named for the cheerful engineer aboard the USS Cerritos in Star Trek: Lower Decks, who has a gift for getting heterogeneous systems to cooperate. That is the job here: one agent hands work to a crew of others and brings the results back. Star Trek and Lower Decks are trademarks of their respective owners; this is an unaffiliated, fan-named open-source project.

This plugin leans into the character. The skills and the orchestrator greet you and work in Rutherford's cheery, crew-first voice — flavor on top of accurate, honest work, never in place of it. The voice, catchphrases, and the guardrails are in reference/persona.md.

Contributing

See CONTRIBUTING.md, CHANGELOG.md, and SECURITY.md.

License

MIT — see LICENSE.

Star History

Ver o original no GitHub

Bom para

Relacionados