convez.ai
AgenteAutomaçãoAvançado
ryan-yuuu

Calfcord

ryan-yuuu/calfcord

Enxames de agentes IA distribuídos colaborando no Discord

Ver no GitHub
BomPluginScore · de 0 a 100

Framework para criar enxames de agentes IA distribuídos que colaboram via Discord. Cada agente tem responsabilidades, ferramentas e memória próprios, podem ser deployados em qualquer host e se comunicam seamlessly uns com os outros através de canais Discord.

Instalação

$npx convez-ai ryan-yuuu/calfcord

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
7
Forks
0
Acompanhando
0
Saúde
100%
Último commit
há 2 semanas
Versão
—
Issues abertas
27
Contribuidores
1

Casos de uso

  • Equipes de agentes autônomos para gestão de tarefas e projetos
  • Automação de workflows com múltiplos especialistas IA colaborando
  • Integração de IA no fluxo de trabalho existente via Discord

Veredito

Calfcord é uma solução sofisticada para orquestração de agentes IA distribuídos com suporte nativo a Discord. Ideal para times que querem automação complexa com múltiplos agentes especializados, mas exige experiência técnica e infraestrutura para aproveitar todo o potencial.

README

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

🐮 Calfcord

CI License: Apache-2.0

Collaborate with your team of AI agents on Discord — each with its own responsibilities and memories, all able to talk to you and to each other.

Distributed by design: agents and tools are independently deployable anywhere. Have agents deployed on your personal laptop, work laptop, and cloud VM all seamlessly collaborate with eachother.

_📸 Demo coming soon

What you get

  • 💬 Communicate with the team on Discord. You and your agent team collaborate on work and provide updates through Discord.
  • 🎭 Agents with distinct job responsibilities and tools. Each agent is a first-class worker with personal responsibilities, tools, and memories.
  • 🌎 Split agents and tools across hosts anywhere in the world. Agents and tools within a team are microservices deployable anywhere, even across hosts.
  • ✏️ Easily onboard new agents to the team. A new agent can be configured in a Markdown file, independently deployed, and added to the team in <2 minutes.
  • 🤝 Agents seamlessly collaborate. Agents chat with each other through private chats, and every exchange is recorded in a Discord thread.
  • 🧠 Bring your own model. Anthropic, OpenAI, other OpenAI-compatible APIs, or use your ChatGPT Plus/Pro subscription (via Codex) — set it per agent.
  • 🛠️ Built-in tools. Agents can get task-tracking, computer filesystem access, and web search tools by default.

Quick start

You'll need a Discord server you own. No Docker required — the installer bootstraps everything calfcord needs to run locally (a native Tansu broker, a single Kafka-compatible binary, plus the process supervisor).

1. Set up the Discord app (~5 min, one time) — follow docs/discord-setup.md. All you need from it is the bot token and application ID; the wizard discovers your server and channel for you.

2. Install. One line, no Python/Docker/git needed first:

curl -fsSL https://raw.githubusercontent.com/ryan-yuuu/calfcord/main/scripts/install.sh | bash

When it finishes, restart your shell (or open a new terminal) so the calfcord command is on your PATH.

3. Run the guided setup. One continuous session takes you all the way to a live agent:

calfcord init

It walks you through picking a provider + model (the model is chosen from a live list fetched from the provider — you can't mistype an invalid slug), naming your agent (default assistant), and choosing its tools (every built-in, all selected by default). Then you paste your Discord token (verified on the spot) and invite the bot — the wizard waits and auto-detects the moment it joins, then discovers your server and channels for you. Finally it opens your workspace, brings your agent online, and watches until it sees the first reply. No separate broker terminal, no juggling processes. Pick Codex for your provider and it logs you in inline via a device code (a URL + one-time code you open on any device, so it works the same locally or over SSH).

4. Say hello. In any channel the bot can see:

@assistant hello

A reply appears from your agent. init has already verified this end-to-end, so this is your confirmation. You're live. 🎉

What you just built

Your workspace — a local message bus and the Discord bridge — now runs in the background. Your agent is a teammate that clocked in. From here you can add more teammates, turn on the receptionist that answers messages nobody @mentioned, or split the team across machines — all without restarting the workspace. The same config and the same commands work whether your org runs on one laptop or twenty hosts.

Day to day

A handful of commands cover everyday use:

calfcord status                 # who's online — the org board
calfcord agent create <name>    # define a new teammate (a Markdown file)
calfcord agent start  <name>    #   ...then bring it online in the live org
calfcord logs -f                # tail the workspace as it runs
calfcord stop                   # close the workspace
calfcord start                  #   ...and reopen it later (substrate only)

calfcord start brings up just the workspace (broker + bridge); teammates (agents, tools, the router) clock in on demand with ... start. The full task-by-task guide is docs/using-calfcord.md.

Going further — I want to…

Pick your next move by goal:

I want to…Go to
Create or customize an agent (fields, models, tools)docs/authoring-agents.md
Give agents more toolsdocs/authoring-tools.md
Let agents answer messages without an @mentiondocs/ambient-routing.md
Use my ChatGPT Plus/Pro subscription (Codex)docs/codex-auth.md
Have agents talk to each other (A2A)docs/a2a-threads.md
Run agents across machines / go to productiondocs/distributed-deployment.md
Understand how it worksdocs/architecture.md (or run calfcord explain topology)
See everything I can do, task by taskdocs/using-calfcord.md
Configure every settingdocs/configuration.md
Review security / threat modeldocs/security.md
Fix something that's brokendocs/troubleshooting.md

Define your own agent

The installer seeds a provider-agnostic starter, assistant, in ~/.calfcord/agents/ — text-only until calfcord init configures it (init bakes in the provider, model, and tools you picked, replacing the pristine seed). Your agents live there and survive calfcord self update. This section is the guide for adding more agents by hand.

An agent is one Markdown file. Drop a new one into ~/.calfcord/agents/:

---
name: scribe
display_name: Scribe
description: Friendly assistant that answers concisely.
avatar_url: https://api.dicebear.com/9.x/glass/png?seed=scribe
provider: openai
model: gpt-5-mini
tools: [private_chat]
thinking_effort: medium
---

You are Scribe, a friendly AI agent. Be helpful and reply concisely (1–3 sentences).

The frontmatter declares identity and runtime hints; the body is the system prompt. The filename must match name, and the slash command is always /<name>. Drop the file in, then bring it online with calfcord agent start scribe.

Prefer not to hand-write the file? The full lifecycle is on the CLI: calfcord agent create | list | show | edit | set | rename | delete (plus calfcord agent tools for just the tool list). create is a guided wizard, edit an interactive field menu, and set its scriptable equivalent. After editing a running agent, apply the change with calfcord agent restart <name>.

Full field reference (providers, models, tool scoping, thinking effort) and the calfcord agent CLI → docs/authoring-agents.md.

How it works

Calfcord runs in two layers, so you can scale the team without touching the plumbing:

  • The workspace (substrate) — the always-on background office: the broker (a local message bus) and the calfkit-bridge (the single Discord gateway). calfcord start brings this up, detached and health-gated; nothing else runs until you ask for it.
  • The roster — teammates that clock in and out of the running workspace. Each maps to one of calfcord's worker process types:
    • calfkit-agent — runs the agent(s). calfcord agent start <name>.
    • calfkit-router — decides who answers un-mentioned ambient messages. calfcord router start.
    • calfkit-tools — runs the tool(s). calfcord tools start.

Every one of these is an independent microservice talking over the broker, so any of them can run on any host. Move the roster onto other machines and point them all at one shared broker URL — same config, same commands, no rewrite. Run calfcord explain topology for the one-screen version, or see docs/architecture.md.

Configuration

calfcord init (from the quick start) writes ~/.calfcord/config/.env with the essentials. To edit settings later, re-run calfcord init or open that file directly; docs/configuration.md is the complete environment-variable reference.

Documentation

Contributing

Python 3.12+, dependencies managed with uv (uv sync, then uv run pytest). See CONTRIBUTING.md, CODE_OF_CONDUCT.md, and SECURITY.md.

License

Apache-2.0.

Ver o original no GitHub

Bom para

Relacionados