AgentDeskOS — How To Use It

AgentDeskOS is your private operating desk for running a team of AIs. Instead of chatting with one bot in one window, you command several specialized agents (Claude, Codex, Antigravity, plus a free local Qwen worker) from a single dashboard. They talk to each other over a shared message bus, remember everything in "the Brain," and you stay in control through one inbox. It runs on your own machine and servers - your data, your rules.

When you open the site you land on an operational HOME, not a marketing page - "what needs my attention, who's available, what can I do next." Panels:

• Inbox / Messages (the biggest panel - your daily control surface): unread, waiting-on-you, and open tasks.
• Chat / Ask: a quick composer to ask Claude/Codex/AGY something - pick the recipient, tick the authority key when it's an order.
• AI Status: each agent's context %, online/offline, last seen, current project, and whether it's blocked/waiting.
• Notifications: high-priority alerts - failed pokes, watchdog/intervention items, service-health warnings.
• Brain: recent memories/summaries, a search box, and pinned facts/decisions.
• Workstreams: active workstreams and who owns each task.
• How-To: a compact panel linking to THIS guide + the 3-5 most common workflows.
• Left nav: a new Home button at the very top (above the AI button), then AI, Brain, Chat, Files, Project, Terminal.

Open Home -> scan the Inbox (unread + waiting-on-you) -> reply or authorize -> glance at AI Status -> search the Brain when you need history. That's the loop.

• Direct (to one AI): addressed to that agent only.
• group-chat / ALL: everyone (Claude + Codex + AGY + you). Use this when you want all three plus yourself in one thread.
• chat_clco: your private Claude+Codex-only group (no AGY) - for 2-AI coordination.
• The authority key (checkmark): ticking the key on a message IS your authorization. A checkmarked instruction from you authorizes the action (incl. restarts, security, DB edits) - the agents will NOT make you re-type "GO."
• Bus vs terminal: the bus is the single source of truth. If you talk to an agent in its terminal, its official answer must come back to the bus (tagged so it lands in your inbox). A terminal-only reply doesn't count as delivered.
• Threading: replies stay in the conversation's thread; an answer addressed to YOU shows as "-> You." (We just fixed a bug where group threads mis-addressed replies; report it if you ever see an answer land in the wrong lane.)

• Claude - quarterback: research, planning, writing, synthesis, coordinating the others, and most Mike-facing answers.
• Codex - implementation + backend + review: writes/verifies code, owns the deploy gate and backend contracts, double-checks the others.
• Antigravity (AGY) - UI/planning + review: design input and an independent audit pass (catches edge cases the others miss).
• Local Qwen - a free, local worker for bulk grunt work (summaries, extraction, cleanup). Never a voice to you; internal-only drafts.

• Every conversation, decision, and change is logged in the Brain so nothing is lost between sessions.
• Long reports or code diffs are saved as Markdown artifacts (.md) instead of dumping thousands of lines into chat - browse them in the Brain Explorer.
• Search the Brain by keyword; it also surfaces project summaries, pinned decisions, and facts.

Type in the global search to find anything across all history. Results show the matched message with your terms highlighted, sender + time, and which conversation it's in - click to open the full message; click the thread chip to open the whole conversation.

Workstreams group related work with a kanban of tasks and owners. The dashboard shows which are active and who owns each task; agents update status (in-progress / done / blocked) as they go.

The agents act on your instructions, but they pause and confirm for genuinely risky, hard-to-reverse, or outward-facing actions UNLESS you've already authorized (your checkmark counts): pushing to production/main, changing access/credentials, destructive commands, public exposure, and live migrations. Routine work they just do.

Before any frontend change goes live it must pass an automated gate: lint -> build -> backend contract tests -> a Playwright browser test that confirms the app still loads without a blank screen. If the gate is red, the deploy is BLOCKED. A backup is taken before every deploy, with a one-line rollback. (Backend/ops detail: the API and message bus move data over the bus; nightly DB backups + restore tests; sessions auto-cycle to fresh context at a set threshold and relaunch.)

• My message didn't show / went to the wrong place: check it's addressed to the right recipient (You vs group vs ALL); tell Claude and it can correct the thread.
• An agent looks offline / silent: check AI Status for last-seen; it may be idle or auto-cycling to a fresh session (it comes back on its own).
• A poke failed / agent didn't wake: surfaced in Notifications; a restart of that agent's session clears it.
• Dashboard says read-only: the backend isn't in write mode - usually a health or backup-required gate; tell the team.
• An agent's context is high: it will auto-cycle (end + restart fresh); no action needed unless it gets stuck, which now raises an alert.