Current version: 0.3.0. See CHANGELOG.md for release history.
Every Claude Code or Codex session becomes an orchestrator. You conduct; cheaper models play.
Conductor makes your session's expensive model do only what expensive reasoning is for - understanding intent, decomposing work, routing, judgment, and synthesis - and delegates the actual labor (writing code, running greps, drafting, reviewing) to cheaper pinned models. Because weekly plan limits weight usage by per-token cost, and execution tokens dwarf orchestration tokens, this cuts your plan burn substantially: usually 60-80% less for build-heavy work (the same fact as stretching a plan 3-5x longer), less if your default model is already a cheap one. Works on Claude Code today; Codex CLI support ships in this release too.
curl -fsSL https://conductorskill.com/install.sh | bashThen open a new session and hand it something real:
/conductor add pagination to the users API and cover it with tests
It plans on your primary model, dispatches the work to sonnet builders and
haiku scouts in parallel, verifies with a critic panel, and reports back with
a one-line usage split so you can see where the tokens went.
Uninstall anytime:
curl -fsSL https://conductorskill.com/install.sh | bash -s -- --uninstallInto ~/.claude/:
| Path | What |
|---|---|
agents/builder.md |
sonnet - all hands-on implementation work |
agents/scout.md |
haiku - cheap recon (greps, inventories, status checks) |
agents/critic.md |
sonnet - adversarial review / verification panels |
agents/architect.md |
opus - deep design for epics (used sparingly) |
skills/conductor/SKILL.md |
the /conductor orchestration loop |
CLAUDE.md |
a marked, idempotent doctrine block that makes every session delegate by default, and every agent write less |
Into ~/.codex/ and ~/.agents/, when Codex CLI is detected:
| Path | What |
|---|---|
~/.codex/AGENTS.md |
a marked, idempotent doctrine block for Codex sessions |
~/.codex/agents/*.toml |
the same crew (builder, scout, critic, architect), pinned to GPT-5.4 / GPT-5.4-mini / GPT-5.5 |
~/.agents/skills/conductor/ |
the orchestration loop for Codex sessions, invoked as $conductor (or from /skills) |
The models (sonnet/haiku/opus on Claude Code; GPT-5.4/GPT-5.4-mini/GPT-5.5
on Codex) are aliases that resolve to whatever your plan provides, so the same
kit works whether your primary model is Opus, GPT-5.5, or anything else you run
as your session default.
- Backs up
CLAUDE.mdand any agent it would replace before touching them. - Writes its doctrine between
<!-- BEGIN CONDUCTOR (managed) -->markers, so re-running updates the block instead of duplicating it, and--uninstallremoves exactly that block and nothing else. - Records a manifest so uninstall cleanly reverses the install.
- Stays interactive under
curl | bash; setCONDUCTOR_YES=1for unattended.
- Claude Code with access to the
sonnetandhaikumodel tiers (any paid plan).opusis used only by thearchitectagent for epic-scale design; if your plan has no Opus, skip that agent - everything else still works. - Codex CLI with GPT-5.4 and GPT-5.4-mini. GPT-5.5 is used only by the
architectagent for epic-scale design. - Copilot support is a fast-follow. Gemini / Antigravity support is deferred pending that product's transition.
- Optional: an improvement skill such as
ebi-processand any domain-expert skills. Conductor uses them if present and skips them if not. - Per-agent model pinning requires a current CLI: Claude Code mid-2026 or later, or Codex CLI with custom agents support.
Two parts:
- Always-on doctrine (the
CLAUDE.mdblock): every session reads it and behaves as a project manager - it delegates instead of doing the labor itself. It also carries a write-less economy ladder so every agent (and the loop's own reports) defaults to terser output - fewer wasted tokens on either side of the delegation. - The
/conductorskill: invoke it to run the full structured loop (recon -> plan -> parallel dispatch -> verify -> synthesize) on a big objective. Routine returns come back as compact columnar JSON with a declared row count; anything touching auth, credentials, money, migrations, deletions, or anything else destructive or irreversible still returns plain JSON, and when a worker is unsure it stays plain.
Everything installed is tracked in ~/.claude/skills/conductor/.install-manifest.
Backups from an install run carry a .bak.<timestamp> suffix and are left in
place after uninstall.
Conductor cutting your burn? Add the badge. It points people back here:
[](https://conductorskill.com)MIT (see LICENSE). Third-party attributions (ideas reimplemented,
no code copied) are in THIRD-PARTY-NOTICES.md.