An AI workforce doesn't keep office hours, so its oversight can't either. The same platform — same identity, same policies, same audit trail — is on the web, on your phone, on the desktop, in the browser toolbar, inside Claude, and on a voice line. Six surfaces. One control plane.
The governance layer holds consequential actions for a human. That promise is only real if the human can act from wherever they are — otherwise held work piles up until someone loosens the gate to make the pain stop.
So every surface carries the queue, the policy context and the trail. The controls stay tight because acting on them stays easy.
The full platform: build agents, teams and AI employees, wire tables and knowledge, write policy, watch runs stream token by token, and read the audit trail one tool call deep. Everything on this site describes what the web app can do — it is the deepest surface, and the one your builders live in.
The mobile app is built around the moments that can't wait for a laptop. The approvals queue with aging chips and grouped asks. A morning briefing of what every worker shipped yesterday. Any run, drilled down to its tool calls with sanitized payloads. Budget forecasts with the projected breach date. And critical alerts — a circuit-breaker halt, a hard cap — that push through quiet hours when you've said they should.
A native desktop app for the people who work alongside their AI coworkers all day: chat with employees and teams, watch live runs, and keep the approvals queue one keystroke away instead of one browser tab among forty. Notifications land at the operating-system level, so a held action surfaces even when the browser is closed.
The Chrome extension brings the coworker to the page you're on. Hand work to an employee from any site, check what's running, and clear approvals without leaving the tool you were in. The gate doesn't require a context switch — which is what keeps it from being bypassed in practice.
The platform is its own MCP server. Connect it to Claude and describe what you need — the client can search the gallery, install packs, configure agents, run work and read results, all through the same permissioned API your users get. AI operating AI, inside the same governance layer.
Agentic teams are voice-capable over a telephony-grade stack: talk to a team, hear it answer, keep the same session, memory and audit trail as the typed conversation. A surface, not a separate product.
