convez.ai
AgenteDesenvolvimento WebIntermediário
pumanitro

AgentsFlow

pumanitro/AgentsFlow

UI desktop para orquestrar agents Claude Code com to-do list e git preview

Ver no GitHub
InicialPluginScore · de 0 a 100

Interface desktop Electron + Next.js para gerenciar agents Claude Code em background. Oferece lista visual de sessões (to-do), terminal embutido, editor Markdown, visualização de árvore git e histórico por diretório.

Instalação

$npx convez-ai pumanitro/AgentsFlow

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
2
Forks
1
Acompanhando
0
Saúde
42%
Último commit
há 2 dias
Versão
v12.0.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Gerenciar múltiplas sessões de agents em background de forma visual
  • Inspecionar mudanças no código enquanto agents trabalham em tempo real
  • Editar e revisar arquivos Markdown inline durante sessões de agentes

Veredito

Ferramenta especializada para quem trabalha intensamente com Claude Code agents. Transforma gerenciamento CLI árido em UX visual e fluida, com file tree colorido por git status e terminal embutido — mas limitada a usuários do Claude Code e macOS.

README

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

AgentsFlow

Home – pinned conversations and tracked directories Embedded terminal attached to a background session File editor with git-aware sidebar

Electron + Next.js desktop UI for Claude Code's background agents — track only the sessions this app launches, treat them like a to-do list, and inspect the working tree of each one without leaving the window.

Status: early. Built against Claude Code CLI v2.1.139+ (claude agents / claude --bg).

Why

Built to scratch three specific itches working with Claude Code day-to-day:

  1. Agent management in the CLI is ugly. claude agents is functional but unpleasant to live in — no real overview, no pinning, no per-directory grouping. AgentsFlow turns background sessions into a visual to-do list scoped to the agents you launched from this app.
  2. No native Markdown editor. When an agent is working in a repo full of .md files, there's nowhere to read or edit them comfortably without alt-tabbing to another app. AgentsFlow ships an Obsidian-style live Markdown editor inline, next to the chat.
  3. No tree preview. Watching agents touch a codebase is hard without seeing the working tree. AgentsFlow shows a git status-colored file tree per session, with an embedded shell, so you can inspect changes the moment the agent finishes — no terminal-juggling.

Highlights

  • Pinned conversations as a to-do list — every spawn is pinned by default; click ✓ Done to archive into the directory's history. ↻ Reopen brings it back.
  • Per-directory history — ⋯ menu on each tracked-directory card opens the full list (pinned + done) for that repo.
  • Live status + auto-titles — title comes from Claude's auto-generated short name (e.g. "joke delivery response"), description streams in from state.json/detail via fs.watch so changes appear within tens of milliseconds.
  • Embedded terminal — clicking a pinned row attaches to that session via claude attach <short> inside an xterm.js panel. ⌘+← (or the Back button) detaches; ⌘+→ from home opens the focused row.
  • File / git sidebar — every session view has a left sidebar with two modes:
    • Changes — git status tree, colored by status (red = untracked, yellow = modified, green = added, etc.)
    • Files — full project tree, with the same git colors layered on, ignored files & dirs faded
    • Expand/collapse state is persisted per conversation.
  • Image paste — paste an image into the spawn input; it's saved to <dir>/.agentsflow/images/ and the absolute path is appended to the prompt so Claude can Read it. Auto-cleaned when the conversation is removed (and orphan-swept on startup).
  • Help modal — ⓘ button in the top right lists all shortcuts and shows the app version.

Requirements

  • Node 18+
  • Claude Code CLI v2.1.139+ on $PATH (claude --version)
  • macOS (the only platform tested so far)

Develop

npm install
npm run rebuild      # rebuild node-pty against Electron headers (one-time)
npm run dev          # next dev + electron, hot-reloads renderer

DevTools auto-opens (detached) in dev mode.

Production build & run

npm run build        # next build (static export) + tsc Electron
npm run start        # launch Electron with the prebuilt renderer

How it works

  • Each dispatched session is launched as claude --bg --permission-mode bypassPermissions <prompt> in the chosen tracked directory; no --name is passed so Claude auto-generates the title.
  • The session's daemonShort is parsed from claude --bg stdout; the full sessionId is resolved by polling claude agents --json (stdout is read via a tempfile to dodge an 8 KB pipe-truncation issue in Electron).
  • Per-session metadata is read directly from ~/.claude/jobs/<short>/state.json — watched with fs.watch for instant updates, with a 30-second fallback tick from claude agents --json to catch missed events.
  • Attaching spawns claude attach <short> via node-pty and pipes it to xterm; detaching kills only the PTY — the background session keeps running.

Storage

A single JSON file under Electron's userData directory.

  • macOS: ~/Library/Application Support/AgentsFlow/store.json (or …/Electron/store.json if the productName isn't set).

Per-conversation tree-expand state is in the renderer's localStorage.

Keyboard

WhereKeysWhat
Home↑ / ↓Move focus between pinned conversations
Home⌘+→ (or Alt+→)Open the focused conversation's terminal
Terminal⌘+← (or Alt+←)Detach and return to home
TerminalShift+EscSame
AnywhereEscClose modal

License

MIT. See LICENSE.

Ver o original no GitHub

Bom para

Relacionados