Welcome to Orchemax

Orchemax is the local runtime that lets several AI coding agents work on your projects without wrecking the repo.

What Orchemax is#

• A single local binary, orch, that runs underneath the agent CLIs you already use — Claude, OpenCode, Codex, Cursor, Copilot and the rest.
• It elects one Chair per project (the first interactive seat) and lets that Chair spawn workers for secondary tasks.
• It puts the workshop's memory on disk — ledger, board, resume, last-session card — instead of in a chat that ends when the window closes.
• It enforces the shared-code law: locks on shared folders, orch:export as the publish contract, and a symbol index every agent can query before it writes a duplicate.
• It can proxy model traffic through a local gateway that holds your provider keys, rotates them, and crushes payloads before they leave the machine.

What Orchemax is not#

• Not a replacement chat or IDE. You still type your prompt inside the agent's own UI. orch claude opens Claude; it does not wrap it in another interface.
• Not a cloud service you have to join. Linking a workspace to the Orchemax overlook is optional; an unlinked workshop keeps working offline.
• Not a model provider. Orchemax ships no LLM of its own. You bring the keys, or the agent brings its own plan login.
• Not a mandate. The seeded doctrine files, the guards and the hooks are opt-in. Your own process wins if you prefer it.

How the pieces fit#

  your terminal
      |
      |  orch claude / orch codex / orch opencode ...
      v
  +--------------------------------------------------+
  |  orch (local runtime)                             |
  |    Chair  --dispatch_worker-->  worker  worker    |
  |    locks | symbols | bus | disk memory | gateway  |
  +--------------------------------------------------+
      |                                    |
      v                                    v
  your workspace on disk            your model providers
  (orch.yaml + .orch/)              (keys stay local)

Team and automation#

Once one seat works, these turn a workshop into something a team can run and plug into the rest of its tooling — every item here shipped with a unit test and a real-agent probe:

• Notifications and approvals — a signed webhook on sealing and promotion, local Telegram soft notify (Professional+), and SaaS Assist HITL from Team+.
• Team work schedules — company timezone and per-day On/Off/Ignored windows that gate Team dispatch in plain language (no cryptic plan deny strings).
• Common seat tune-ups — terminal UI Orchemax owns, vendor MCP/tool allowlists, account/plan login seats (no API key), OpenClaude model pinning, dontAsk vs bypassPermissions, Notification wire.
• System tray presence — tray icon that opens orch ui, glances gateway/workers/HITL, and bridges desktop toasts.
• Know when a worker finishes or asks — a desktop toast with a sound per event, Claude Code's own stop hook, and a Telegram /answer when a worker's question needs a human.
• Inbound webhooks — a local, HMAC-signed receiver n8n or Zapier can call to create and dispatch a task.
• Cross-vendor review — one CLI writes, a different vendor's CLI reviews read-only before promote.
• Principles Orchemax seeds, and how to add yours — the nine-line default layer, which lines a gate actually checks, and your own layer on top.
• Silent seat protocol vs workshop files — SessionStart / .orch/ORCH_RUNTIME.md injection vs files you edit; no public */skills/orch-* by default (ORCH_LEGACY_SKILLS=1 escape).
• Same CLI, many projects — --name + to_project when several Chairs share one agent binary.
• Add your language — a seed under .orch/langs/ so the index reads a language this build does not ship, and the PR that gives it to everybody.
• Your own rules and Packs, audit trail and policy — gates and skills you copy in, edit, and can audit after the fact.
• Evidence bundle — one artifact per session: verdict, files, commands, tests, tokens — attachable to a PR.
• Resource-aware scheduling — queue a dispatch instead of piling onto a loaded host; watch it in orch status.
• Clean re-setup — inventory and remove what Orchemax or an agent left in a workspace, re-seed from a known state, or orch uninstall Orchemax entirely.
• Plan mode and worker gateway keys are covered in Chair and workers and Gateway and overlook; full-text memory search in Disk memory; daily token budgets in Show my usage and my leaks; AGENTS.md as the one instruction file every CLI reads in Configure agents.

Where to go next#

• Install — copy to %LOCALAPPDATA%\Programs\orch (Windows) or ~/.local/bin (macOS/Linux), then PATH + machine home.
• Quickstart — bind a workspace, seed it, open the first seat, run the first worker.
• What to read next — the path that matches what you came here to do.
• Release 0.7.0 — preview, early access by invitation: a nightly rehearsal, seal gates scoped to what a change reaches, lessons on probation, test adapters per language, a publish guard that fails closed.
• Release 0.6.0 — preview, early access by invitation: a session start with a real brief, a close that saves while you work, a live progress line, quieter notifications.
• Release 0.5.0 — preview, early access by invitation: gates at the moment of action, the improvement cycle, a self-trimming footprint.
• Release 0.4.0 — what is new: core lessons, fast seals, orch land, capacity scheduler, one binary.