Claude Ultracode Agents
payne0420/claude-ultracode-agents
Delegue tarefas de código para Codex, Cursor e opencode via Claude Code
Ver no GitHubConjunto de skills para Claude Code que delega trabalhos de codificação a agentes IA externos (Codex, Cursor, opencode) com orquestração de workflows. Permite rodar múltiplos backends de código em paralelo ou sequência, com suporte a revisão, planejamento e implementação headless.
Instalação
npx convez-ai payne0420/claude-ultracode-agents1 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
Casos de uso
- Revisão de código com múltiplos modelos especializados
- Orquestração de workflows que alternam entre agentes Claude e externos
- Implementação headless de features com isolamento em git-worktree
- Chamadas diretas a endpoints OpenAI/Anthropic compatíveis
Veredito

Excelente para times que usam múltiplos agentes de código e precisam orquestração avançada. Requer configuração local cuidadosa (credenciais fora do git) e CLIs dos backends instalados — não é plug-and-play, mas poderoso para fluxos de trabalho complexos.
README
A tradução para o português deste README ainda está na fila.
Agent Skills for Claude Code
A small bundle of Claude Code skills for delegating coding work to external AI coding agents — OpenAI Codex, Cursor, and opencode — plus a raw OpenAI-/Anthropic-compatible HTTP endpoint backend, and an orchestration layer that lets ultracode workflows run any of them instead of (or alongside) regular Claude subagents.
What's inside
| Skill | What it does |
|---|---|
ultracode-external-agents | Author ultracode Workflow scripts whose agent() steps run on Codex / Cursor / opencode instead of Claude. Explains the bridge, ships copy-paste delegation helpers, a runnable multi-model review-panel example, and per-task "which backend(s)" guidance. |
codex-exec | Drive OpenAI's Codex CLI non-interactively (codex exec) — second opinions, code review, planning, headless implementation, structured output. |
cursor-agent | Drive Cursor's CLI agent headless (cursor-agent -p) — reviews, planning, implementation, git-worktree isolation. |
opencode | Drive the model-agnostic opencode CLI (opencode run) — reviews, planning, implementation across many providers. |
llm-endpoint | Call a configured OpenAI-compatible (/v1/chat/completions) or Anthropic-compatible (/v1/messages) endpoint — a raw model call, not an agent. Base URL + key are read from local config, never committed. |
ultracode-external-agents builds on the other three and points back to them
(via the Skill tool) for the full per-CLI reference.
Install
Skills live in ~/.claude/skills/. Copy the folders you want there:
git clone https://github.com/payne0420/claude-ultracode-agents.git
cd claude-ultracode-agents
cp -R ultracode-external-agents codex-exec cursor-agent opencode llm-endpoint ~/.claude/skills/
Start a new Claude Code session and the skills will be available. Invoke them by intent (e.g. "review this with codex", "get a second opinion from opencode") or by name.
Requirements
- Claude Code (the agent harness).
- For each backend you want to use, the matching CLI installed and
authenticated:
codex,cursor-agent, and/oropencode. Each skill checks availability before relying on the tool. ultracode-external-agentsis meant for an ultracode / Workflow authoring turn (Claude Code's multi-agent Workflow tool must be active).
Notes
- Model defaults are configurable and machine-specific. The helpers default
to codex
gpt-5.5@xhigheffort, cursorcomposer-2.5, and opencodeopencode-go/deepseek-v4-pro@--variant max. opencode model ids in particular vary per machine — runopencode modelsand repoint as needed. Every default is overridable per call. - These skills delegate work to external / third-party models. Don't put secrets in prompts you send to them.
llm-endpointneeds local config (kept out of git). Copyllm-endpoint/config.example.shto~/.config/llm-endpoint/env,chmod 600 ~/.config/llm-endpoint/env, and fill in your base URL, API key, and default model. That file is the only place the endpoint URL/key live — they are never committed to this repo.
License
MIT.