Concepts

How It Works

The flow from local AI activity to reviewed workspace knowledge, approved actions, and audit history.

Review flow

Governed proposal flow

Agents submit proposals locally. The app and workspace policy decide what becomes shared company state.

Producer

Agent, connector, or script finds useful signal

mos CLI

Local proposal without platform credentials

Menu bar queue

Review item with refs, policy, and actions

Human decision

Approve, deny, reroute, or request changes

Workspace sync

Wiki write, action, record, or audit receipt

Credentials stay with the app and managed workspace. Agents submit through mos and do not bypass local review.

Agents propose. The menu bar reviews. Approved work becomes shared workspace state.

Codex / Claude / Cursor / Hermes / Granola / local scripts
  -> MOS skill or adapter
  -> mos event submit --file -
  -> menu bar review queue
  -> approve, deny, reroute, or request changes
  -> workspace record, wiki write, action, or audit receipt

The important boundary is that agents emit structured proposals. The user and workspace policy decide whether those proposals become shared company state.

System steps

1. Producer: An agent, local app, connector, scheduled job, or script notices something worth reporting.
2. Local event: The producer writes event JSON through mos. The proposal is local first and does not require the agent to hold workspace credentials.
3. Review item: The menu bar app shows what happened, why it matters, source references, privacy posture, and available actions.
4. Decision: A human approves, denies, reroutes, snoozes, requests changes, or opens the dashboard for deeper review.
5. Workspace sync: Approved proposals become workspace records, wiki updates, source actions, or audit receipts according to policy.

Surfaces

Menu bar: The local attention and approval surface. It is optimized for small decisions, routing, and awareness.
Dashboard: The workspace control surface for agents, sources, wiki, approvals, governance, runs, and history.
CLI: The scriptable local interface that agents and skills use to validate and submit proposals.
Workspace wiki: The durable company knowledge layer that approved wiki writes update.

Evidence over vibes

Governed AI work needs deterministic evidence where possible: source references, diffs, tests, run logs, policy decisions, and commit history. Multiplayer OS keeps those artifacts attached to the proposal so review is grounded in more than an agent summary.