Frugal Orchestrator
thxmxx/frugal-orchestrator
Orquestra modelos baratos; Opus só valida e verifica
Ver no GitHubSkill que implementa orquestração inteligente de modelos: o modelo caro (Opus) apenas orquestra, valida e verifica, enquanto subagentes explicitamente fixados usam modelos mais baratos (Sonnet, Haiku) para geração pesada. Reduz consumo de tokens caros em ~60-80% em sessões multi-step.
Instalação
npx convez-ai thxmxx/frugal-orchestrator1 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
- Sessões longas com múltiplos passos que queimam quotas caras
- Revisões de código grandes e leitura profunda de bases de código
- Pesquisa, análise e processamento de dados pesados
- Migrações de código e drafting de relatórios extensos
Veredito

Ferramenta pragmática e bem-pensada para quem enfrenta limites reais de orçamento de API. Resolve invisibilidade do inherit de modelos caros, mas requer compreensão de quando ativar orquestração — não é plug-and-play puro. Impacto real em sessões pesadas de multi-step.
README
A tradução para o português deste README ainda está na fila.
frugal-orchestrator
Weekly caps and API spend are the bottleneck — not capability.
A Claude Code skill that encodes one discipline: when your session runs on an expensive model, the expensive model only orchestrates, gates, and verifies. All heavy generation is delegated to explicitly model-pinned cheap subagents.
Why
Every Agent tool and Workflow tool subagent silently inherits the main-loop model unless you pass an explicit model: parameter. If your main loop is Opus, every subagent is Opus. Nothing warns you — output looks fine, the cap burns invisibly.
A heavy multi-step session can exhaust a weekly Opus cap in an afternoon. With this skill:
| Without frugal-orchestrator | With frugal-orchestrator |
|---|---|
| Every subagent inherits Opus | Opus only on decompose + verify gates |
| 5–10× expensive-bucket burn rate | Sonnet on deep reads, drafts, research |
| Cap exhausted mid-session | Haiku on grep, extraction, classification |
Result: ~60–80% reduction in expensive-model token spend (illustrative; calibrate against your own session logs). The numbers are order-of-magnitude estimates from the author's heavy multi-step sessions. Your actual reduction depends on task structure and session length.
Install
claude plugin marketplace add thxmxx/frugal-orchestrator
claude plugin install frugal-orchestrator@frugal-orchestrator
Verify
claude plugin list
# frugal-orchestrator@frugal-orchestrator enabled
claude plugin details frugal-orchestrator@frugal-orchestrator
# Shows: always-on tokens, on-invoke tokens, skill: frugal-orchestrator
Update
claude plugin marketplace update frugal-orchestrator
claude plugin update frugal-orchestrator@frugal-orchestrator
Remove
claude plugin uninstall frugal-orchestrator@frugal-orchestrator
claude plugin marketplace remove frugal-orchestrator
Then sweep for any leftover hooks or statusLine entries (see plugin-hygiene if you want a skill that automates this cleanup). If you ever used /frugal-orchestrator:off, also remove the leftover flag file: rm -f ~/.claude/frugal-orchestrator.off.
Usage
Auto-load (default): A SessionStart hook injects the full skill into every session (startup, resume, and /clear). The injected directives make the discipline active by default:
- Heavy or multi-step work — large code reviews, large codebase reading, multi-file builds, deep research, migrations, long report drafting — is decomposed and delegated to model-pinned cheap subagents.
- Workflow-by-default: substantive multi-step work is orchestrated via the Workflow tool (installing the plugin is your standing opt-in to multi-agent orchestration), degrading gracefully to parallel model-pinned Agent calls when Workflow is unavailable.
- Trivial single-step tasks and conversational turns stay inline — no orchestration overhead.
Opt out / back in:
/frugal-orchestrator:off # disables auto-load (creates ~/.claude/frugal-orchestrator.off, persists across sessions)
/frugal-orchestrator:on # re-enables auto-load (removes the flag file)
While disabled, the skill remains available for explicit invocation — ask Claude to "apply frugal-orchestrator" or "use model-pinning discipline" on any task.
Context cost: the auto-injected skill adds roughly 1k tokens of always-on context per session — that is the price of never silently inheriting the expensive model.
Before / After Example
Task: Assess a 3000-line legacy module for migration to a new API.
Without frugal-orchestrator:
- Main loop (Opus) reads the full module inline
- Dispatches analysis subagents — all inherit Opus
- Each sub-analysis burns Opus tokens at full rate
- Cap exhaustion risk: high
With frugal-orchestrator:
- Opus decomposes: "extract deprecated-API usages (Haiku), analyze each for breaking-change risk (Sonnet), adversarial review (Haiku), synthesize and gate (Opus)"
- Haiku extracts usage list → returns JSON (cheap, mechanical)
- Sonnet analyzes each usage with file:line citations → returns structured findings (capable, mid-cost)
- Haiku adversarially refutes findings → returns objections list (cheap cross-check)
- Opus reads only the objections list and final findings JSON, decides which modules are confirmed ready (expensive model on gate-only work)
- Estimated result: ~60–80% fewer Opus tokens for equivalent analysis quality
Origin
This discipline originated from the author's "Opus-minimal" operating model for heavy AI-augmented analysis sessions, where the expensive-model weekly cap is the literal binding constraint on throughput. The pattern generalizes to any heavy multi-step problem: large code reviews, large builds, deep research, migrations, and long report drafting.
Sibling project: plugin-hygiene — inspect installed Claude Code plugins by real token cost and remove them cleanly.
License
MIT © 2026 Lucas Thom