convez.ai
AgenteDevOps & InfraestruturaIntermediário
yucc-leon

AgentBoard

yucc-leon/AgentBoard

Dashboard para agentes Claude/Codex — monitore e responda de qualquer navegador

Ver no GitHub
InicialPluginScore · de 0 a 100

Interface web para gerenciar sessões de agentes Claude e Codex rodando em tmux, localmente ou via SSH. Permite monitorar, responder e retomar conversas de agentes de qualquer dispositivo, com resumos inteligentes das sessões.

Instalação

$npx convez-ai yucc-leon/AgentBoard

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
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Monitorar agentes autônomos em execução contínua
  • Retomar conversas com agentes de semanas anteriores
  • Controlar agents em máquinas remotas via SSH
  • Acessar status de agentes do celular enquanto está fora

Veredito

Solução prática para quem roda agentes em background e precisa acompanhá-los remotamente. A arquitetura minimalista (sem banco de dados, baseada em tmux) é robusta, mas requer conforto com terminal e infraestrutura de rede.

README

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

AgentBoard 🧠

English · 中文

I leave Codex and Claude Code running on my Mac and walk away. AgentBoard is how I check in from my phone: see what the agents are doing, type a reply, or pull up something I ran last week and keep going. That's really all it's for.

Dashboard

What it is

A little web app in front of your agent sessions. A session is just a tmux pane running an agent CLI, so AgentBoard lists the ones you have open (on this machine, or on boxes you reach over SSH) next to the Codex and Claude conversations already sitting on disk. Open any of them and start typing.

If you've added an LLM key, it also gives each conversation a short title and a recap of where things stand, what's next, and the loose ends you probably forgot, so you don't have to re-read a wall of text to remember what you were doing.

Recovery card

How it works

It leans on three tmux commands and not much else:

  • It finds running agents with tmux list-panes (or ssh <host> tmux list-panes), and your past conversations by reading the JSONL logs Codex and Claude already write under ~/.codex and ~/.claude. No database, and nothing to install on the remote machines.
  • It types into a pane with send-keys and shows you the output with capture-pane. The terminal tab gets a real pty stream.
  • When you open an old conversation and hit send, it brings that conversation back up in tmux and delivers your message. You don't click "resume" first.

Expose it to the internet and one token guards everything.

Quickstart

uv sync
uv run agentboard init        # writes ~/.agentboard/config.yaml
uv run agentboard web         # local hub at http://127.0.0.1:8765

Agents already running in tmux show up on their own. To start a fresh one, hit + New.

Getting to it from your phone

uv run agentboard web --remote

This binds publicly and prints a token, the access URLs, and a QR code. Scan the QR with your phone and you're in. The token sticks around as a cookie for 30 days, so you only do this once per device. Lost the token? agentboard token prints it again; agentboard token --rotate gives you a new one. How you expose the port is up to you: Tailscale, cloudflared, an SSH reverse tunnel, whatever.

A heads-up on speed: it's snappy on the same Wi-Fi and slower from a different network, especially if your traffic goes through a relay. That lag is only in the controls. The agent itself keeps working on your machine at full speed.

CLI

CommandWhat it does
agentboard initCreate ~/.agentboard/config.yaml
agentboard sessionsList agent sessions across machines
agentboard send <machine> <name> <msg…>Type a message into a session
agentboard new <machine> <cwd> [--command codex] [--name x]Start a session
agentboard kill <machine> <name>Kill a session
agentboard summarize [-m machine] [-n name]Build LLM summary cards
agentboard token [--rotate]Print the access token + URLs + QR (or rotate it)
agentboard web [--port 8765] [--remote]Start the web hub

Configuration

~/.agentboard/config.yaml:

workspace:
  data_dir: ~/.agentboard

machines:
  - name: local
    type: local
    codex_home: ~/.codex
    claude_home: ~/.claude
    tmux: true
  - name: h200
    type: ssh
    host: h200          # must work as `ssh h200` (use ~/.ssh/config)
    codex_home: ~/.codex
    claude_home: ~/.claude
    tmux: true

llm:                    # optional: only for titles & summaries
  base_url: https://api.deepseek.com
  model: deepseek-v4-flash
  api_key_env: DEEPSEEK_API_KEY

remote:
  enabled: false        # `web --remote` flips this on
  bind_host: "0.0.0.0"

Without an LLM key the titles are just the first line of your opening message; add one and they get written properly.

Privacy

Everything stays on your machine under ~/.agentboard/. A conversation only goes to an LLM when you ask for a title or summary, and secrets (API keys, tokens, private keys) are stripped out first. Remote access is off until you turn it on, and gated by the token when you do.

Development

uv sync --extra dev
uv run --extra dev pytest
uv run --extra dev ruff check

Contributing

This started as a personal tool, so the rough edges you run into are genuinely useful to hear about. Issues, PRs, and stars are all welcome.

Acknowledgements

The interactive side (tmux-first sessions driven from a web hub) took cues from StarAgent.

License

MIT

Ver o original no GitHub

Bom para

Relacionados