Agent Chat Skill
akostibas/agent-chat-skill
Messaging entre agentes Claude Code na mesma máquina, sem infraestrutura
Ver no GitHubSkill para Claude Code que permite múltiplas sessões de agentes na mesma máquina trocarem mensagens em tempo real via log JSONL. Funciona como Slack para agentes Claude, com notificações automáticas push e sem infraestrutura adicional.
Instalação
npx convez-ai akostibas/agent-chat-skill1 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
- Frontend e backend agents sincronizando mudanças de contrato durante desenvolvimento
- Múltiplos agentes em worktrees diferentes reportando bugs e conflitos
- Agente coordenador distribuindo trabalho e coletando resultados de workers
Veredito

Solução elegante e minimalista para orquestração local de múltiplos agentes Claude. Resolve um problema real de coordenação sem deps pesadas — ideal para workflows complexos em máquina única, mas com escopo limitado ao ambiente local.
README
A tradução para o português deste README ainda está na fila.
agent-chat
A Claude Code skill that lets two (or more) Claude Code sessions running on the same machine exchange notes mid-task. It's like Slack, for Claudes.
When it's useful
- One agent iterating on a frontend while another works on the backend it calls — they flag contract changes to each other as they happen.
- Two agents on different worktrees of the same repo, reporting bugs and merge conflicts across the fence.
- A coordinator agent fanning work out to several worker agents and collecting results on one channel.
- A documentation agent in contact with the implementor agents, keeping docs current as multiple agents land feature work.
Features
- Push, not poll — peer messages arrive in chat as notifications automatically. Agents don't loop or burn turns waiting.
- Broadcast by default,
@nameto narrow — everyone on the channel sees a message unless you address it to specific peers. - Spoof-proof — every message is prefixed with its true sender, so one agent can't impersonate another.
- Zero infrastructure — pure shell over a per-user JSONL log. No daemon, no network, no binary to build.
- Self-cleaning — idle channels are pruned automatically.
Install
From a local clone:
make install
This copies skill/ to ~/.claude/skills/agent-chat/ and marks the scripts executable. Override the destination with SKILL_DIR=... — e.g. make install SKILL_DIR=/path/to/project/.claude/skills/agent-chat for a project-level install.
The skill invokes its own scripts via ${CLAUDE_SKILL_DIR} (the directory its SKILL.md was loaded from), so it runs correctly from either a user-level (~/.claude) or project-level (<project>/.claude) install. Channel state, however, is always shared under $HOME/.claude/agent-chat/ regardless of where the scripts live — that's what lets agents in different directories talk on the same machine (override with AGENT_CHAT_ROOT=... if you ever need isolated channels).
Then add these entries to the permissions.allow array in the matching settings.json so subagents (and the auto-permission classifier) don't block the scripts. For a home install that's ~/.claude/settings.json; for a project install use the project's .claude/settings.json and its install path:
"Bash(bash ~/.claude/skills/agent-chat/join.sh:*)",
"Bash(bash ~/.claude/skills/agent-chat/send.sh:*)",
"Bash(bash ~/.claude/skills/agent-chat/history.sh:*)",
"Bash(bash ~/.claude/skills/agent-chat/stream.sh:*)"
Requires jq and shlock on PATH. shlock ships with macOS at /usr/bin/shlock. jq is one brew install jq away.
Upgrading from agent-mail
The skill was previously named agent-mail. If you have the old install, remove it and migrate any in-flight channels:
rm -rf ~/.claude/skills/agent-mail
mv ~/.claude/agent-mail ~/.claude/agent-chat 2>/dev/null || true
Then update the four permissions.allow entries in ~/.claude/settings.json from agent-mail to agent-chat.
Use
In each Claude Code session, tell Claude:
Join agent-chat on channel
<slug>.
Claude will pick a name describing what it's working on, and join a "chat room". Other Claude sessions may join and send messages to the channel. Messages arrive to Claude agents in the channel as notifications automatically.
See skill/SKILL.md for the full instructions Claude follows.
Limitations
- Local machine only — no cross-host channels.
- Designed for full Claude Code sessions, not subagents. Subagents have constrained turn lifetimes and tend to end their turn before notifications can drive a reply.
- Channel dirs whose
loghasn't been touched inAGENT_CHAT_TTL_DAYSdays (default 14) are pruned opportunistically each time any of the skill scripts runs. Override by exportingAGENT_CHAT_TTL_DAYSin your shell.
License
MIT.