Cross-Agent Skill Bridge
anombyte93/cross-agent-skill-bridge
Compartilhe skills entre agentes de IA sem duplicação ou conflitos
Ver no GitHubFerramenta para compartilhar skills (habilidades) entre diferentes agentes de código (Claude Code, Codex, OpenCode) usando um hub neutro em ~/.agents/skills. Sincroniza seletivamente, sem sobrescrever modificações locais, mantendo uma única fonte de verdade.
Instalação
npx convez-ai anombyte93/cross-agent-skill-bridge1 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
- Manter skills sincronizadas entre Claude Code, Codex e OpenCode
- Detectar skills divergentes e resolvê-los antes de sincronizar
- Criar fonte de verdade centralizada para skills reutilizáveis
Veredito

Ferramenta bem pensada para quem trabalha com múltiplos agentes de código. Resolve um problema real (sincronização segura de skills) com abordagem não-destrutiva e transparente. Documentação clara, instalação simples, sem dependências externas — produção-ready para workflows multi-agent.
README
A tradução para o português deste README ainda está na fila.
cross-agent-skill-bridge
Author an agent skill once, use it in every coding agent you run.
Different coding agents read "skills" (a SKILL.md directory) from different places:
Claude Code reads ~/.claude/skills, OpenCode reads ~/.config/opencode/skills,
and Codex reads the shared ~/.agents/skills via native skill discovery. So a skill
you write for one agent is invisible to the others until you copy it over — and doing that
by hand lets the copies quietly drift apart.
This is a tiny, dependency-free tool that treats ~/.agents/skills as a hub and
each agent as a spoke, and syncs skills between them — selectively (you choose
what to share) and non-destructively (a copy you've edited is never silently
overwritten).
Why I built it
I maintain a set of skills across Claude Code and Codex. I wrote one
(repo-sanitize) in Claude and wanted
it in Codex too — but I didn't want to blanket-copy everything, because some skills
are intentionally agent-specific and some had drifted. I needed to see the spread
first, then share just the ones I meant to. That visibility + safe-sync is the whole
tool.
Install
One command (any agent):
curl -fsSL https://raw.githubusercontent.com/anombyte93/cross-agent-skill-bridge/main/install.sh | bash
Clones the repo, symlinks the skill into the neutral ~/.agents/skills hub, and runs a
smoke test so you can see it working immediately. Idempotent and non-destructive — re-run
to update; it never deletes your skills. Requires git and python3 (3.9+).
Claude Code (plugin marketplace):
/plugin marketplace add anombyte93/cross-agent-skill-bridge
/plugin install cross-agent-skill-bridge
Use
cd skills/cross-agent-skill-bridge/scripts
# 1. See what's where, and what's diverged:
python3 skill_bridge.py status
# 2. Preview a share (changes nothing):
python3 skill_bridge.py share --from claude --skill my-skill --dry-run
# 3. Do it (copy to all other present agents):
python3 skill_bridge.py share --from claude --skill my-skill
# Keep a single source-of-truth live everywhere instead of copying:
python3 skill_bridge.py share --from claude --skill my-skill --mode link
status prints a matrix: rows are skills, columns are agents, cells are content
hashes (link:… = symlink, - = absent). A ⚠ DIVERGED row means the copies differ
— decide which is canonical before syncing.
Safety
- Non-destructive: a diverged target is skipped with a warning unless you pass
--force. - Source is read-only: bridging never deletes or edits the source skill.
- In-sync = no-op: identical copies aren't rewritten.
--dry-runto preview any operation.
Extending to other agents
export CROSS_AGENT_SKILL_ROOTS="codex=~/.codex/skills,work=~/work/.agents/skills"
Scope
Skills only. MCP servers and agent definitions are agent-specific config and out of scope for v1.
Credits
See CREDITS.md. ~/.agents/skills is the directory Codex reads via native
skill discovery, and the symlink-install pattern is prior art from the agent-skills
ecosystem — notably Superpowers by Jesse Vincent
(MIT). This tool does not invent that mechanism; it just makes living on both sides of it
painless. (Path scope is agent-specific — Claude Code uses its plugin marketplace, OpenCode
uses ~/.config/opencode/skills/; see CREDITS.)
License
MIT.