Consilium - Orquestrador Multi-Agent para Coding
TemurTurayev/consilium
Coordena Claude, Codex e Gemini para deliberação e revisão cruzada sem queimar quotas
Ver no GitHubOrquestrador que coordena múltiplos agentes de IA (Claude Code, Codex CLI, Gemini CLI) para deliberar coletivamente sobre problemas complexos, revisar trabalho cruzado e distribuir tarefas respeitando os limites de quota de cada subscrição. Implementado em Rust, funciona com as CLIs oficiais autenticadas, sem chamar APIs diretamente.
Instalação
npx convez-ai TemurTurayev/consilium1 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
- Deliberação multi-agent sobre problemas técnicos complexos (revisão anônima e síntese)
- Auditoria de diffs de código com múltiplos revisores e saída de exit codes
- Otimização de quota distribuindo tarefas pesadas entre provedores com limites mais livres
- Accounting preciso de tokens respeitando semântica diferente de cada provedor
Veredito

Projeto sofisticado que resolve um problema real de economia de quota em setups multi-provedor. A arquitetura é sólida (M1 e M2a completos, E2E verificado), mas ainda em desenvolvimento (M2b e M3 pendentes). Para equipes que realmente usam múltiplas subscrições de IA e precisam de orquestração inteligente.
README
A tradução para o português deste README ainda está na fila.
Consilium
Get a second opinion. And a third.
One orchestrator for the AI coding subscriptions you already pay for. Consilium drives the official CLI agents — Claude Code, Codex CLI, Gemini CLI — so they can deliberate on hard problems, cross-review each other's work, and split tasks between providers without burning through any single subscription's limits.
Named after the medical consilium: specialists from different fields gathering around one patient.
Why
- Subscriptions, not API keys. Consilium never calls provider APIs directly — it runs the official CLIs authenticated with your existing Claude Max / ChatGPT / Google plans. All their features keep working; nothing gray-zone.
- Quota economics changed. Since June 15, 2026, Anthropic meters
claude -p(headless) against a separate monthly credit instead of the flat subscription pool. Consilium's architecture is built around this: heavy lifting goes to the workers with the freest quota, and the planned attached mode (M3) keeps the conductor inside your interactive Claude Code session via MCP, where flat subscription limits still apply. - Accurate accounting is the headline feature. Provider token semantics genuinely differ — we verified each against recorded real CLI output (see table below). Most tools get at least one of them wrong.
Status
| Milestone | Scope | State |
|---|---|---|
| M1 — Engine foundation | CLI adapters, session manager, quota store, doctor/run/quota commands | ✅ Done — verified E2E |
| M2a — Deliberation | council (anonymized peer review → chairman synthesis), review (diff audit with CI exit codes) | ✅ Done — 72 tests, verified on live providers |
| M2b — Execution | conduct (conductor/workers), auto pipeline, supervisor, quota-aware routing | 🚧 Next |
| M3 — Server & UI | axum + WebSocket server, MCP attached mode, React web UI, quota dashboards | Planned |
| v1.1+ | Warp terminal integration (OSC 777), Tauri desktop app | Planned |
Quick start
Prerequisites: Rust ≥ 1.85 and at least one of the agent CLIs installed and logged in (claude, codex, gemini).
git clone https://github.com/TemurTurayev/consilium.git
cd consilium
cargo build --release
cargo run -q -- doctor # check agent CLIs
cargo run -q -- run --provider gemini "Reply with: ok" # single-agent smoke run
cargo run -q -- quota # usage in the last 5h window
# The flagship: convene a council — workers answer independently, cross-review
# each other anonymously, the chairman synthesizes the best answer.
cargo run -q -- council "Async Rust: when is spawning a task per request wrong?"
# Audit a diff with the reviewer role. Exit codes: 0 no critical findings,
# 2 critical findings, 3 reviewer output unparseable (fails closed).
git diff | cargo run -q -- review --diff-file /dev/stdin
Every deliberation writes a full JSON transcript to ~/.consilium/runs/ — including
the anonymization map and per-reviewer scores, so you can audit who said what
(and whether an agent favored its own anonymized answer).
$ cargo run -q -- doctor
✓ claude 2.1.111 (Claude Code)
✓ codex codex-cli 0.139.0
✓ gemini 0.36.0
How it works (M1)
claude / codex / gemini CLIs (your subscriptions)
│ headless stdout
▼
adapters/ pure parsers: raw CLI output → normalized AgentEvent
│ (fixture-tested against RECORDED REAL outputs)
▼
sessions.rs tokio process spawn, event streaming, stderr-drain
│
▼
AgentEvent SessionStarted · Message · Thinking · ToolCall ·
│ FileChanged · Usage · Completed · Failed
▼
quota.rs SQLite usage log, sliding-window aggregation
Token semantics actually differ per provider
Verified against recorded real CLI outputs (core/tests/fixtures/*/recorded/):
| Provider | Input side | Output side |
|---|---|---|
| Claude | input + cache_creation + cache_read — cache tokens are disjoint additions | output |
| Codex | input only — cached_input_tokens is a subset, summing would double-count | output + reasoning_output_tokens |
| Gemini | Σ over all internal models (prompt + cached) — one request may use several | Σ candidates + thoughts |
Development
- TDD throughout;
cargo test(unit + integration),cargo clippy --all-targets -- -D warnings,cargo fmt. - Adapter parsers are tested against recorded real CLI outputs committed as fixtures — format regressions surface without spending quota. Re-record with
script/record_fixtures.sh(spends a few real requests). - Design spec:
docs/specs/· implementation plans:docs/plans/
Roadmap highlights
council— the llm-council pattern ported to coding agents on subscriptions: independent answers → anonymized cross-review → chairman synthesis.auto(default mode) — council for planning, conductor for execution, cross-review per subtask, supervisor watching everything ("reads a lot, writes rarely" — input tokens are cheap).- Attached conductor — your interactive Claude Code session orchestrates workers through Consilium's MCP server, spending subscription limits instead of the programmatic credit.