Orchestrator

A self-hosted Windows desktop app that runs an LLM Director which chats with you and spawns specialised agents (pm / researcher / coder / qa / devops / security) to do the work.

Built on top of the claude (and optionally codex) CLIs you already have installed — Orchestrator drives them; it doesn't bundle them.

Single-user, self-hosted, Windows-only for now. v0.6.0.

What it does

• Director chat in one pane: describe the task, get back a plan, accept it, watch agents run.
• Six agent roles with role-specific system prompts, tool allow-lists, and per-role skill slots:
  • pm — read-only, decomposes work
  • researcher — read + web fetch, writes findings to disk
  • coder — read + edit + shell, implements
  • qa — read + edit + shell, writes and runs tests
  • devops — read + shell, builds and deploys
  • security — read + shell + web, audits for vulnerabilities
• Live drawer per agent: KPI strip (Step / Tokens / Cost / Elapsed), tools usage, context tab, system prompt, pause.
• Sequential auto-spawn so the Director runs pm → coder → qa in order, not all at once.
• Per-agent budgets in dollars, tokens, or wall-clock seconds. Hard-stop on exceed.
• Session persistence — Director chat, agents, log lines all rehydrate on restart.
• Spend screen with real per-day cost chart and a deep link to claude.ai for official usage.
• Two providers — Claude (default) or Codex (codex exec --json). Each project picks one at creation.
• Auto-update via update-electron-app against update.electronjs.org (Microsoft-operated, TLS-pinned to GitHub Releases) every 10 minutes.

Install

Grab the latest Orchestrator-Setup.exe from the Releases page and run it. SmartScreen will warn — the build isn't code-signed.

Once installed, Orchestrator updates itself in-place. A pill appears in the top bar when a new version finishes downloading.

You'll also need at least one of:

• Claude — claude CLI installed and logged in (Anthropic API key, Claude Code OAuth, or Claude Team plan login all work).
• Codex — codex CLI installed and logged in. ChatGPT-plan users are supported; gpt-5-codex is the default model.

Orchestrator detects which CLIs are present and gates project creation accordingly.

Develop

git clone https://github.com/lukeeexd/Orchestrator.git
cd Orchestrator
npm install
npm start

Stack: Electron Forge + Vite + React 19 + TypeScript + sql.js (WASM SQLite). The main process holds the runtime; the renderer is a three-pane shell (Director / Agents / Drawer). Branch model is dev → main; tags vX.Y.Z on main trigger the release workflow.

To produce a Windows installer:

npm run make

Output lands at out/make/squirrel.windows/x64/Orchestrator-Setup.exe.

Architecture in one breath

The main process spawns claude (or codex) as a child process per agent and streams the structured event log back over typed IPC. The renderer holds a useAgents hook that mirrors that stream into the UI. The Director itself is just another long-running session — its outputs include a fenced orchestrator-plan JSON block that the UI parses into the plan card.

For the longer story (decisions locked, milestones, deferred features), see PLAN.md. For UI work, see docs/design/README.md.

Status & contributions

This is a single-developer project. Issues are welcome; PRs are accepted on a case-by-case basis. The roadmap in PLAN.md is the source of truth for what's coming next.

License

MIT.