Cockpit — Centro de Comando com Claude Agent
ArthaudMe/cockpit
Centro de comando desktop com Claude para integração com tools corporativas
Ver no GitHubCockpit é um co-pilot desktop que funciona como centro de comando corporativo, conectando-se a ferramentas em tempo real (Gmail, Slack, Linear, GitHub, Notion, Google Calendar) e permitindo fazer perguntas, obter contexto e tomar ações via Claude Code Agent. Oferece sistema multi-agente, 7 conectores de dados, memória persistente e interface em estilo terminal com suporte a tabelas, gráficos e blocos Mermaid.
Instalação
npx convez-ai ArthaudMe/cockpit1 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
- Consolidar contexto de múltiplas ferramentas em uma interface única
- Automatizar consultas e ações em email, calendário, issues e mensagens
- Criar agentes especializados sob demanda para tarefas específicas
- Manter histórico e memória de conversas com injeção de contexto automática
Veredito

Solução ambiciosa e bem arquitetada para agregação de contexto corporativo via Claude. Prototipo funcional com 7 integrações reais e sistema de agentes robusto, mas em fase experimental (Mio harness). Recomendado para equipes que precisam centralizar múltiplas ferramentas e construir workflows baseados em agentes.
README
A tradução para o português deste README ainda está na fila.
Cockpit
Command center for your company. Desktop AI co-pilot that connects to your live tools and lets you ask questions, get context, and take action — all from one place.
What's live
AI Engine Layer
- Claude-first, multi-backend — Claude is the primary engine; Codex and Ollama available as alternates in the model switcher
- Backend auto-detection — scans installed binaries and versions at startup
- Multi-agent system — instant agent creation ("+" tab), double-click to rename, per-agent backend/model switching
- One chat engine — main chat and focus-view chat share the same agents, memory, and history
- Warm process pre-spawning — per-agent warm CLI processes for instant first tokens; system prompts rebuilt on every respawn with fresh datasource context
- Conversation memory — recent turns and relevant workspace history travel with each message (one-shot CLI calls, stateful conversations)
- Streaming response parsing — handles incomplete JSON blocks mid-stream without breaking
- Skills system — 12 built-in skills with slash commands, per-skill prompt injection, and enable/disable toggles; LLM can propose new custom skills mid-chat
- Subagent spawning — LLM suggests specialized subagents; user approves via inline button, subagent opens as a new tab
Data Integrations (7 connectors + MCP)
- Google Calendar — next 7 days of events, auto token refresh with 5-min expiry buffer
- Gmail — recent emails with unread status
- Linear — assigned issues (excl. canceled/completed), priority mapping via GraphQL
- GitHub — open PRs involving user, notifications
- Notion — recent pages by last edit, OAuth + internal token fallback
- Slack — recent channel messages (past 24h), parallel fetch with username caching
- Granola — meeting notes from local macOS cache (last 7 days), mtime-cached parsing
- Generic MCP client — connect any MCP server (stdio or SSE) as a datasource; resources auto-injected into AI context
Context & Intelligence
- Dynamic system prompt builder — user profile + live calendar, Slack, Linear, GitHub, email, Notion, Granola, and MCP data assembled per spawn
- Knowledge layer — datasource history persisted daily to
~/.cockpit/history/; keyword search over past items injected as historical context with each message - Memory — two bounded markdown files (MEMORY.md + USER.md); LLM-driven add/replace/remove via
cockpit_memoryblocks, injection-pattern scanning, live entries in every prompt - Render blocks — tables, bar charts, card grids, layouts, and mermaid blocks embedded inline in responses via
cockpit_renderJSON - Context Focus — click any calendar event, feed item, metric, or search result to open an entity-focused chat with structured data and suggested questions
- Project inference — projects auto-clustered from Linear/GitHub/Slack signals via LLM, with heuristic fallback; clustering cached on disk and re-run only when sources change
Dashboard & UI
- Live activity feed — color-coded by type, built from all connected sources
- Context sidebar panels — calendar (grouped by day), Slack highlights, skills, todos
- Chat interface — terminal-style input, markdown rendering, image attachments (paste/drop), slash-command autocomplete
- Live refresh — Electron main process polls every 60s and pushes to the renderer over IPC (browser dev mode falls back to polling)
Auth & Token Management
- OAuth via token-exchange proxy — client secrets never ship in the app or repo; they live in a small Vercel function (
proxy/) - Auto token refresh — 5-minute buffer before expiry, transparent to caller
- Secure local storage — tokens in
~/.cockpit/with0o600permissions, mtime-cached reads - Deep-link OAuth callback —
cockpit://protocol forwards to the local server
Settings & Onboarding
- 2-step onboarding — Claude install/sign-in (auto-detecting, with one-click install) → connect one tool, add the rest later
- Settings — profile, connected tools, AI engines, analytics opt-in; MCP servers, skill toggles, custom skills, and agent management under Advanced
- Persisted state — debounced localStorage writes, capped per-agent chat history, images excluded from persistence
Search
- Command palette (Cmd+K) — instant client-side search across cached data, then live API search with LIVE badge
- Live search providers — Gmail, Google Calendar, Linear, GitHub, Notion, Slack
- Source filtering —
in:slack,in:linear,in:github, etc.
Write-Back Actions
- Action cards — LLM proposes actions via
cockpit_actionJSON; user reviews and approves inline - 6 action types —
linear_create_issue,github_comment_pr,slack_send_message,calendar_create_event,gmail_draft,notion_update_page - Action log — executed actions persisted to
~/.cockpit/action-log.json(capped at 500)
Analytics & Crash Reporting
- PostHog integration — opt-in anonymous usage analytics (off by default); no personal data or chat content sent
- React error boundary + global error listeners; Electron process-level handlers log to
~/.cockpit/crash-log.json - Renderer crash recovery — auto-reload on
render-process-gone/unresponsive
Desktop App
- Electron shell — Mac (DMG), Windows (NSIS), Linux (AppImage) targets
- Standalone server bundle — the app ships Next.js standalone output (self-contained
server.js), notnode_modules; runs on a dynamically allocated localhost port - Auto-update —
electron-updaterwith GitHub Releases, 4-hour check interval (requires a signed app on macOS) - Tray icon, window state persistence, splash screen, graceful shutdown
- Background intelligence — periodic rule checks over live data with native notifications
Prerequisites
- Node.js 20+
- pnpm
- Claude CLI installed and authenticated (
claudecommand available in your terminal)
Optional backends: Codex, Ollama (detected automatically).
Setup
cp .env.example .env.local # fill in OAUTH_PROXY_SECRET (see below)
pnpm install
pnpm dev
Open http://localhost:3939. The onboarding flow guides you through the rest.
OAUTH_PROXY_SECRET authenticates the app to the OAuth token-exchange proxy and is inlined at build time — it must match PROXY_SECRET on the deployed proxy (see proxy/README.md). Without it, datasource OAuth connects will fail.
Connecting datasources
Connect services from onboarding or Settings. You'll be redirected to each service's OAuth page; tokens are stored locally in ~/.cockpit/tokens.json. Granola needs no OAuth — its local cache is detected automatically.
Desktop (Electron)
pnpm electron:dev # development (uses the dev server on :3939)
pnpm electron:build # next build + standalone packaging + electron-builder
pnpm electron:publish # build and publish a release (auto-update feed)
Quality
pnpm typecheck
pnpm test
Architecture
src/
app/
api/ # API routes (agents, chat, datasources, projects, skills, ...)
page.tsx # Main dashboard (IPC-pushed data, browser polling fallback)
components/
columns/ # ProjectsColumn, FeedColumn, ChatColumn, ContextColumn
layout/ # Header, NotificationBell
renderers/ # Table, BarChart, CardGrid, Layout, Mermaid render blocks
ui/ # ChatMessage, CommandPalette, action/skill cards
views/ # ContextualChatView, OnboardingView, SettingsView
lib/
agent-manager.ts # Multi-agent lifecycle, warm processes, persistence
agent-stream.ts # Shared CLI→HTTP streaming (one chat engine)
prompt-prelude.ts # Per-message context: recent turns + historical items
context.ts # System prompt builder with live datasource context
fs-cache.ts # mtime-keyed JSON read cache for ~/.cockpit stores
datasources/ # Connectors, manager (single-flight cache), token store
knowledge/ # History writer/search + conversation persistence
projects/ # Project store + LLM-based project inference
skills*.ts # Built-in defs, custom skills, extraction
electron/
main.js # Spawns the standalone server, tray, auto-update, IPC push
proxy/
api/oauth/token.ts # Vercel function holding OAuth client secrets
scripts/
prepare-standalone.js # Copies static assets into the standalone bundle
Data storage
All user data stays local:
~/.cockpit/tokens.json— OAuth tokens (mode 0o600)~/.cockpit/profile.json— name, role, company~/.cockpit/agents.json— agent configurations~/.cockpit/skills.json/custom-skills/— skill toggles and user-created skills~/.cockpit/mcp-servers.json— MCP server configurations (mode 0o600)~/.cockpit/memories/— MEMORY.md + USER.md (mode 0o600)~/.cockpit/history/— daily datasource snapshots + conversation history~/.cockpit/cache/— offline datasource cache, project-inference cache~/.cockpit/action-log.json— executed action history~/.cockpit/window-state.json,crash-log.json— Electron statelocalStorage— chat UI state (capped, images excluded)
Installing Unsigned DMGs (macOS)
Until the app is signed and notarized, macOS Gatekeeper will block it. After dragging Cockpit to Applications, either run:
xattr -cr /Applications/Cockpit.app
or go to System Settings → Privacy & Security and click Open Anyway after the first blocked launch attempt. Note that auto-update does not work on unsigned macOS builds.
License
MIT — see LICENSE.