Capability Orchestrator
DZMing/capability-orchestrator
Roteamento inteligente de capacidades com detecção de ambiente em tempo real
Ver no GitHubMCP server que escaneia o ambiente local, detecta skills, comandos, plugins, agents e outros MCP servers disponíveis, e roteia automaticamente prompts do usuário para a melhor superfície de execução. Inclui camada de Intent Router que expande prompts operacionais abreviados em contratos de execução estruturados.
Instalação
npx convez-ai DZMing/capability-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
- Automatizar roteamento de tarefas para o MCP mais adequado
- Expandir prompts operacionais em contratos de execução estruturados
- Gerenciar confirmações de segurança para ações críticas (deploy, delete, produção)
Veredito

Ferramenta muito técnica para ambientes Claude Code/Codex que necessitam roteamento sofisticado de capabilities. Excelente para fluxos de trabalho complexos com múltiplas superfícies de execução, mas requer arquitetura bem definida e melhor documentação em português.
README
A tradução para o português deste README ainda está na fila.
capability-orchestrator
Capability awareness and auto-routing for Claude Code and Codex, with an experimental Hermes host adapter.
capability-orchestrator scans the local agent environment, summarizes available
skills, commands, plugins, agents, and MCP servers, then routes user prompts to
the best available execution surface. It also includes a separate Intent Router
layer that expands short operational prompts into a five-part execution contract.
What It Does
- Injects a capability summary when a new Claude Code / Codex session starts.
- Routes matching prompts to the right skill, command, or MCP server.
- Expands shorthand prompts like "continue", "execute", and "what is left" into a full What / Guardrails / Success / Budget / Verify contract.
- Requires confirmation before publish, push, deploy, delete, paid, credential, production, and real product or UX decisions.
- Supports Claude Code and Codex as the stable primary hosts.
- Provides an experimental but verified Hermes host bridge.
- Keeps OpenClaw limited to read-only scan compatibility; the OpenClaw host bridge install path is currently frozen.
- Keeps install, reinstall, uninstall, lifecycle, and release checks executable.
Intent Router
The Intent Router layer is for shorthand operational prompts, not for replacing direct capability matching. It first classifies the prompt and runs a prompt-level safety precheck. It reads live work context, route history, and preferences only when the prompt is a supported shorthand intent or a high-risk action that must fail closed.
Typical intents include:
| Short prompt | Intent | Result |
|---|---|---|
继续 | continue_work | Continue the current safe technical work from live context. |
执行吧 | execute_plan | Execute an already-discussed plan with verification. |
还有什么没做完 | work_status | Summarize the remaining work and pick the next feasible task. |
做到可以商用 | commercial_readiness | Move the project toward a commercially usable, release-gated state. |
The execution contract always includes:
WhatGuardrailsSuccessBudgetVerify
It folds in bounded live context from the repo rules, git status, recent route
log entries, and an optional preference profile at
~/.config/capability-orchestrator/preferences.json. Preferences are advisory
only and never reduce risk. If the prompt does not look like a safe operational
intent, the existing skill / command / MCP matcher still handles direct routing.
Quick Start
curl -fsSL https://raw.githubusercontent.com/DZMing/capability-orchestrator/master/install.sh | bash
Then restart Claude Code or Codex.
Windows Claude Code native install:
powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1
Uninstall:
bash ~/.claude/plugins/cache/capability-orchestrator/install.sh --uninstall
For Codex, replace ~/.claude with ~/.codex.
Host Support
| Host | Status | Notes |
|---|---|---|
| Claude Code | Stable | Uses SessionStart and UserPromptSubmit hooks |
| Codex | Stable | Linux/macOS native; Windows via WSL2 |
| OpenClaw | Frozen, scan-only | May read local workspace skills; no host bridge install, adapter commands, or lifecycle commitment |
| Hermes | Experimental, verified | Runtime snapshot, route bridge, slash command bridge, pre_llm_call bridge, lifecycle verification |
Hermes has verified install/reinstall/uninstall and bridge behavior, but it remains experimental until broader host lifecycle and Windows-native support commitments are frozen. OpenClaw host-bridge support is intentionally frozen; only read-only local skill scanning is kept as compatibility surface.
Advanced Install
# Install a specific release
CAPABILITY_INSTALL_REF=vX.Y.Z \
curl -fsSL https://raw.githubusercontent.com/DZMing/capability-orchestrator/master/install.sh | bash
# Install from master
curl -fsSL https://raw.githubusercontent.com/DZMing/capability-orchestrator/master/install.sh | bash -s -- --channel=master
# Explicit host selection
curl -fsSL https://raw.githubusercontent.com/DZMing/capability-orchestrator/master/install.sh | bash -s -- --platform=codex
curl -fsSL https://raw.githubusercontent.com/DZMing/capability-orchestrator/master/install.sh | bash -s -- --platform=hermes
Verification
npm test
npm run test:all
bash tests/install.test.sh
bash tests/install-idempotent.test.sh
npm run verify:scenarios
npm run verify:host:hermes
npm run verify:host:lifecycle
npm run verify:release
npm run verify:release:strict # required only for an actual release/tag publish
Useful manual checks:
node ~/.claude/plugins/cache/capability-orchestrator/scripts/scan-environment.cjs --mode=awareness
printf '%s' '{"prompt":"show all available capabilities","cwd":"."}' \
| CLAUDE_USER_DIR="$HOME/.claude" \
node ~/.claude/plugins/cache/capability-orchestrator/scripts/route-matcher.cjs --explain
node --test tests/intent-classifier.test.cjs tests/intent-router.test.cjs \
tests/safety-gate.test.cjs tests/prompt-composer.test.cjs \
tests/work-context.test.cjs tests/preference-profile.test.cjs
Safety Model
- The installer updates only capability-orchestrator-owned hook entries.
- Unrelated hooks are preserved during install, reinstall, and uninstall.
- Runtime scans are best-effort and fault-open.
- The scanner does not execute scanned plugin directories.
- MCP servers, plugin manifests, legacy command bodies, and scanned descriptions are advisory matching signals only. They are not executed or treated as instructions.
verify:releaseis a pre-landing audit: it validates package, manifests, supported adapter versions, changelog, tag metadata, GitHub Release state, and rejects any leftover OpenClaw host bridge surface or script.verify:scenariosruns a cross-host Claude/Codex scenario matrix for short prompts, high-risk confirmation gates, skill routing, MCP advisory routing, legacy command safety, and preference redaction.- The route corpus test adds precision/recall-style coverage for shorthand prompts, escape handling, high-risk prompts, skill/command/MCP matches, and no-match behavior.
verify:release:strictis the hard release gate for real publishing; it also requires a clean worktree andHEADmatching the latest release tag.- High-risk intents such as publish, push, deploy, delete, paid actions, credential-gated actions, production changes, and real product or UX decisions require confirmation before action.
- Route logs keep only whitelisted, anonymized fields such as prompt type, target type, confidence, host/source/scope, and MCP trust signals; they do not store raw prompts or credentials.
Documentation
Known Boundaries
- Native Windows support is only committed for Claude Code.
- Codex on Windows should use WSL2.
- OpenClaw host bridge support is frozen; only read-only scan compatibility is retained.
- Hermes is a verified experimental host bridge, not yet part of a formal cross-platform support matrix.
- The Intent Router layer is separate from the direct capability matcher; the matcher still handles skill, command, and MCP routing when the prompt is not a safe operational intent.
License
MIT