agentic-plugins
each4all/agentic-plugins
Framework para colaboração cross-host entre agentes Claude e Codex
Ver no GitHubFramework de colaboração entre agentes de IA em múltiplos hosts, conectando Claude Code e OpenAI Codex como pares. Plugins são escritos uma vez e executados nativamente em qualquer host, com pontes bidirecionais que permitem invocação cruzada entre agentes.
Instalação
npx convez-ai each4all/agentic-plugins1 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
- Orquestração de múltiplos agentes de IA em paralelo
- Desenvolvimento de plugins que rodam nativamente em Claude Code e Codex
- Automação de fluxos de trabalho complexos com ensemble de agentes
- Planejamento estratégico e engenharia de software colaborativa
Veredito

Framework robusto e bem-arquitetado para coordenação de agentes multi-host, com stage-gating claro e documentação técnica madura. Exige expertise avançada em agentes de IA e está em evolução ativa (Stage 3+), recomendado para arquitetos de sistemas e desenvolvedores que precisam de orquestração real entre múltiplas plataformas.
README
A tradução para o português deste README ainda está na fila.
agentic-plugins
Cross-host AI agent collaboration framework.
agentic-plugins bridges Anthropic's Claude Code and OpenAI's Codex CLI as peer agents. Plugins authored once run natively in either host, and each host can invoke the other as a peer agent through bidirectional companion bridges owned by agentic-plugins.
Status
Stage 1: companion layer + first reference plugin shipped. The
architecture and foundational decisions are captured in
docs/adr/ and the overall design in
docs/ARCHITECTURE.md. Stage roadmap lives in
docs/DEVELOPMENT.md.
Available now
The companions/ layer is wire-spec complete (v0.1.1) with both
bridges implemented and tested:
companions/contract.md— wire contractcompanions/claude-companion.mjs— Codex → Claude bridgecompanions/codex-companion.mjs— Claude → Codex bridge
Five installable plugins ship in this repository:
plugins/companions/— script-only library plugin that bundles the canonical companion CLIs for cache-glob discovery by consumer plugins (per ADR-0008)plugins/engineer/— Stage 2 L3 persona plugin: 6 universal cognitive verbs (investigate / frame / decide / compose / critique / refine) with bidirectional companion ensemble. The investigate verb'scited-briefprofile produces durable cited research artifacts (absorbing the Stage 1plugins/researchcontract per ADR-0014)plugins/founder/— second L3 persona plugin for new-business planning per ADR-0036. Incubating scaffold: no user-facing commands yet; the verb surface lands through the ADR-0036 implementation roadmapplugins/orchestrator/— Stage 3+ L2 capability plugin per ADR-0018 §sub-decision-1 and ADR-0019: macro plan, Plan-verify opposite-host peer ensemble, same-host engineer dispatch, manual completion backup, finalize/abort lifecycle, meta continuity commands, and macro auto-archive.plugins/runtime/— Stage 3+ L1 framework primitive per ADR-0024:runtime:doctorreadiness diagnostics, dry-run/defaultruntime:settings, explicitruntime:consensuscompanion execution artifacts,runtime:compathost-version drift planning, read-onlyruntime:worktreeplanning,runtime:contexthandoff/check artifacts, read-onlyruntime:cutoverreadiness evidence, workflow-storage migration, and the pointer-only completion footer.
See each plugin's README for install commands, invocation, and environment details.
The earlier Stage 1 plugins/research reference plugin was retired
at Stage 2.5+ per
ADR-0014; its
cited-brief contract is now folded into engineer:investigate.
Current follow-ups
- Runtime artifact retention/deletion and richer consensus cancellation policy
- Host-native config apply beyond agentic-plugins-owned config
- Risk/budget-driven consensus peer selection and automated synthesis policy
- Non-interactive auth story for CI smoke tests (DEVELOPMENT.md Risk #4)
Concepts
- Hexagonal architecture — host-neutral CORE + per-host ADAPTER + bidirectional COMPANION
- Bidirectional companions —
claude-companion(Codex → Claude) andcodex-companion(Claude → Codex), both owned by agentic-plugins - Standards-aligned core — Agent Skills for skills, Model Context Protocol for tools
- Native install — each host uses its own plugin manager; agentic-plugins does not unify install UX
For consumers
# Claude Code
claude plugin marketplace add each4all/agentic-plugins
claude plugin install companions@agentic-plugins
claude plugin install engineer@agentic-plugins
claude plugin install orchestrator@agentic-plugins
claude plugin install runtime@agentic-plugins
# OpenAI Codex CLI
codex plugin marketplace add each4all/agentic-plugins
# then enable plugins in ~/.codex/config.toml — see plugins/companions/README.md
Each plugin's README documents its invocation surface and environment variables.
For developers
AGENTS.md— primary development guidance (cross-tool standard)CLAUDE.md— Claude Code reference into AGENTS.mddocs/ARCHITECTURE.md— overall designdocs/DEVELOPMENT.md— how agentic-plugins is itself developed (dogfooding plan)docs/adr/— architecture decision records
Relationship to omcc
agentic-plugins is the dual-host successor to omcc.
omcc remains operational (Claude-only) until agentic-plugins reaches feature
parity. See docs/adr/0007-migration-cutover-plan.md
for the cutover plan and
docs/assurance/omcc-cutover-scorecard.md
for the current requirement-to-evidence scorecard.
License
MIT.