Use your harness. To orchestrate many others.
Codex · Claude Code · Copilot CLI · Grok Build · Pi
Orchestrator is a simple skill built on a powerful local CLI inspired by Kubectl. Work with one of your agents; it orchestrates others (Claude Code, Codex, etc) in the background.
You speak in models and outcomes:
Use Fable for the UI, GPT-5.6 Sol for the implementation, and Grok 4.5 for the small fixes. Run independent work in parallel, then have Opus review the result.
Explore the repository with Kimi 2.7. Give the plan to GPT-5.6 Sol, then use Fable only where the interface needs real design work.
Install
npx skills add backnotprop/orchestrator
npm install -g @backnotprop/orchestrator-cli
The CLI install is optional up front. The skill can install it on first use.
/orchestrator launch a gpt5.6 sol agent to build the backend,
a fable agent to design the frontend wireframes,
and after that a grok4.5 agent to build the frontendPreferences
Preferences are optional. Edit
skills/orchestrator/PREFERENCES.md, or
tell your agent what to put there:
Use Fable for extensive UI work.
Use GPT-5.6 Sol for deep execution.
Use Grok 4.5 for small coding tasks.
Use Kimi 2.7 for exploration.
When Fable is usage is capped, use GPT-5.6 Sol.
When GPT-5.6 Sol is unavailable, use Opus.
When every preferred model is unavailable, pause and notify me.
Orchestrator discovers current model names and IDs from installed runtimes instead of relying on stale model slugs.
Human model names are fine. The skill resolves them against live runtime catalogs, checks available provider limits when supported, and follows the current request before saved preferences.
Under The Hood
Orchestrator is not a harness. You use it in your preferred harness and it orchestrates others in the background.
Every worker is a named task with status, logs, output, follow-up, resume, and stop controls. Agents use a small JSON command loop:
orchestrator doctor --json --compact orchestrator models <runtime> --json --compact orchestrator launch <runtime> --json --compact --brief "<task>" orchestrator ps --json --compact --active --brief orchestrator read <task-id>... --wait --json --compact orchestrator interrupt <task-id> --json --compact --reason "<reason>"
The CLI owns process supervision and task state. The calling agent owns judgment, delegation, and synthesis.
Read the Operator Guide for the full control contract, or see custom agents, architecture decisions
License
Business Source License 1.1 today, with production use allowed except for commercial hosted or managed agent-orchestration services. Converts to Apache License 2.0 on July 9, 2029.