The problem: your AI coding agents each live in their own window. Claude Code writes the code, another tool drafts the proposal, a third does research — and you are the glue: copying context between them, remembering where work stopped, and chasing every loose end. A one-person agency drowns in that glue work.
The solution in today's radar: Orkas-AI/Orkas is an open-source desktop app that puts a "Commander" agent in charge of a team. You give it a goal; it plans the steps, delegates each part to the right specialist agent (16 built-in), and can even drive the coding agents you already use — Claude Code, Codex, OpenCode — as local sessions under one plan. Projects hold the goal, rules, memory and files so work continues across days; to-dos and "Auto-advance" keep it moving; finished work always comes back for your review before anything leaves your machine.
Format note: single-featured-factory layout (standing focus set Oct 25, 2026); the Ethiopia business-opportunity section remains removed. Prior reports were available through Nov 2 (yesterday featured coleam00/Archon; Orkas has not been featured before). All facts below verified today from the repo's README, LICENSE, release and commit history via GitHub.
run.cmd: Node 22.12+; workspace at <smallest non-system drive>:\.orkas\. Voice transcription downloads a pinned whisper.cpp runtime on first launch. Python 3 and a C/C++ toolchain only needed for rare native-module source-build fallbacks (npm run native:repair if needed).Modules, artifacts and behaviors verified from the repo README today.
| Agency OS stage | Factory module | Artifact it produces | Which gate consumes it |
|---|---|---|---|
| LEAD | — partial: goals enter as project briefs | Project with goals & rules | Commander routes work inside the project |
| DISCOVERY / PROBLEM_DEFINITION | DeepResearcher — evidence-grounded research, verified sources, flagged contradictions; projects & to-dos | Auditable research report; to-do list split between human and agents | To-dos assigned per agent/human; nothing advances without your review |
| PROPOSAL | ContentWriter, PptMaker, OfficeWorker | Client-ready articles, decks, DOCX/PPTX/XLSX/PDF | Human review of the deliverable before sending |
| SPECIFICATION / ARCHITECTURE | UIDesigner — HTML-first UI from PRDs, goals, screenshots or Figma; ProductDeveloper — repo-aware PRD implementation | UI designs; implementation plans and code | ProductDeveloper claims "verifiable results" — evidence quality untested by us |
| DEVELOPMENT | ProductDeveloper + external CLI agents (Claude Code, Codex, OpenCode) run as local sessions under one Commander plan | Code changes, bug fixes, refactors | Finished work returns for your review |
| QA | — no dedicated test stage (ProductDeveloper "review" only) | — | — (gap) |
| SECURITY | — none (no secret scanning / dependency auditing) | — | — (gap) |
| CEO_REVIEW | Ask-before-send connectors — "Orkas asks before it sends, posts or deletes anything"; finished work returns for review | Approval decisions | Maps to your human-approval rule |
| RELEASE | — no deployment/publish stage (local apps only; publishing a web app needs a paid plan) | — | — (gap) |
| MAINTENANCE | Scheduled automations (hourly→monthly; missed runs catch up within 12h from v1.9) + background learning reviews (agents refine their own skills/memory when idle, subject to cooldowns and execution limits) | Recurring reports; learned skills | Bounded by execution limits and idle-time cooldowns |
| REUSABLE_COMPONENT | Agent marketplace + custom agents — 30+ agents; describe a role and the Commander builds it; per-agent memory and skills | Custom agents, learned skills | Partial — agent-level reuse, not business components |