Claude Code Orchestrator
visiquate/cco
Controle de custos, delegação de agentes e compressão de contexto para Claude Code
Ver no GitHubSistema multi-agent local que rastreia gastos com tokens em tempo real, oferece 117 agentes especializados para delegar tarefas e comprime contexto em até 90%. Roda completamente offline no seu machine, sem enviar dados para a nuvem.
Instalação
npx convez-ai visiquate/cco1 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
- Rastrear e otimizar gastos com API Claude
- Delegar tarefas a agentes especializados (testes, review, docs, segurança)
- Monitorar taxa de acerto do prompt caching
- Reduzir tokens em workflows de IA através de compressão automática
- Implementar limites de orçamento em pipelines CI/CD
Veredito

Solução sólida e prática para quem usa Claude Code em escala. O rastreamento local de custos + compressão automática endereça um problema real (surpresa com bills), mas depende de estar integrado ao seu workflow específico. Recomendado para desenvolvedores com uso consistente e preocupação com otimização de custos.
README
A tradução para o português deste README ainda está na fila.
Hey there, welcome to Claude Orchestra!
If you're here, you probably code with Claude Code and have ever wondered: How much is this costing me? or Is my context efficient? or Could I automate more of this work with specialized agents? If any of that rings a bell, you're in the right place.
What is CCO?
Claude Orchestra (CCO) is your local sidekick for Claude Code. It's a small command-line tool that runs quietly in the background and does three things really well:
- Tracks your AI spending in real time — Every token you use is logged locally (on your machine, never uploaded unless you opt in). You get instant visibility into what things cost.
- Gives you a fleet of 117 specialist agents — Rather than doing everything yourself at the top tier, CCO lets you delegate work to cheaper, focused agents that excel at specific tasks (documentation, code review, security, testing, etc.). Plus a live dashboard to manage them.
- Shrinks your token bills quietly — CCO bundles compression tools (RTK, lean-ctx, Headroom) that cut token use 60–90% on typical tasks. You see the savings tracked in real time.
All of this runs locally. Your data, your machine, your control.
Why you'd want this
Surprise AI bills: Claude Code sessions add up fast, and without visibility, you don't know where the money's going. CCO shows you exactly.
No cache insight: Prompt caching can save a fortune, but you don't know if it's working. CCO shows you hit rates and estimated savings per model tier.
Manual delegation: If you're doing all the coding yourself with Opus, you're paying premium rates for tasks that could be handled by a specialist. CCO's delegation nudge reminds you to use the right tool for the job.
Token bloat: Bash output, logs, and verbose responses eat tokens. CCO's compression tools trim the noise automatically.
Getting started (one command)
curl -fsSL https://raw.githubusercontent.com/visiquate/cco/main/install.sh | bash
That's it. The binary lands in ~/.local/bin/cco and updates itself with cco update.
Next, run the health check (it'll fix things automatically):
cco doctor --fix
This verifies Claude Code is installed, starts the daemon, and optionally sets up compression tools.
Day-to-day commands
Here are the handful of commands you'll actually use:
cco — Launch Claude Code through CCO. Everything gets tracked automatically.
cco tui — The live dashboard. Shows your spend, cache hit rate, active agents, and more. Press ? for keybindings.
cco cost dashboard — Text summary: total spend, breakdown by model, trends.
cco cost cache --period today — Cache efficiency: how much you saved with prompt caching.
cco cost gate --max-usd 10.00 — Budget guard for CI/CD: exits with code 1 if you're over budget.
cco doctor --fix — Health check. Repairs issues automatically.
What to expect
It runs quietly in the background. Once you start the daemon (cco daemon start), it stays out of your way.
Cost tracking is instant and offline. No cloud services, no data uploaded (unless you opt in for telemetry). Everything lives in a local DuckDB file.
First run might be slower. The daemon loads a small semantic embedding model for code search and context compression. After that, it's fast.
Privacy by default. Prompts, responses, and code never leave your machine. Aggregate telemetry (token counts and costs — no text) is opt-out. Transcript upload is off by default and must be explicitly turned on.
Cool stuff to explore
The TUI cockpit (cco tui) is your command center. Five pages: Dashboard (spend + sparklines), Activity (live event tail), Cache (efficiency), Delegation (agent work distribution), and Agents (spawn/kill agents with a form). Press p to filter by project, t to change time range, n to spawn an agent. Full guide at docs/tui.md.
Cache buster detection (cco cost cache) finds cache writes that were never read — wasted money. This is hard to spot manually but CCO finds it automatically.
Token compression — RTK is included by default and cuts Bash output 60–90%. You can also opt into lean-ctx (semantic compression) or Headroom (API-boundary caching). See docs/integrations.md for the full tour.
Delegation nudges — When you're doing implementation work inline with Opus, the system nudges you to use a Sonnet or Haiku specialist instead. Non-blocking, rate-limited, and it learns from your preferences.
MCP server — Register CCO as an MCP server in Claude Desktop or Cursor. Other AI tools can then query your costs, agents, and knowledge graph without the daemon being online. See docs/mcp-and-control-plane.md.
Tune which models your agents use — By default, CCO runs the heavy thinkers on bigger models and the routine work on faster, cheaper ones. Want a different balance? Set one line in your config:
agent_tier = "premium" # options: economy | balanced (default) | premium | max
economy keeps costs lowest; max puts the biggest models everywhere. You can also fine-tune per role (authors vs reviewers vs the lead architect) or per project. Plus, agent teamwork (agents handing off to each other) is on by default.
Multi-tool tracking — cco codex launcher lets you use Codex for code review with CCO's cost tracking. All the savings and metrics flow back into your dashboard.
Getting help
cco doctor — Diagnoses problems and suggests fixes. Usually solves issues automatically with --fix.
docs/ folder — Detailed guides for TUI usage, integrations, cost metrics, and telemetry.
GitHub Issues — Found a bug? Have a feature idea? GitHub issues for the main repo: https://github.com/langstons/cco/issues
You've got this
You're now in control of your AI coding spend, have a toolkit for smarter delegation, and can watch your token compression happen in real time. No surprise bills. No guessing. Just you, Claude, and visibility.
Questions? Start with cco doctor. Want to go deeper? Read docs/tui.md and docs/integrations.md.
Happy coding.
License
MIT — see LICENSE.