AI agent teams for marketing agencies. They plan, run and report on client work. You approve anything that touches a client's account.
A brief in, a report out. The team works between those two points, and you decide the parts that cost money.
One task per client job: a campaign refresh, a monthly report, a landing page. A manager agent splits it and hands the parts out.
Each run sees only its own workspace. No server files, no credentials, no direct internet. Everything they fetch goes through a logged proxy.
A budget change or a campaign edit waits on a card until you click approve. Agents cannot approve their own work.
Cost per task and per client, with a CSV you can attach to an invoice. Agents keep their memory from one job to the next.
A hosted control plane for agent teams, hardened for client work and run by us.
Destructive actions on ad platforms stop at a review card with the exact change. One click runs it. Four hours later it expires untouched.
Every model call is attributed to an agent, a task and a client. The Costs page shows the highest-cost tasks and exports the period as a file.
Runs are isolated from the server and from each other. The outbound proxy records which sites were used and blocks known data drops.
An agent keeps one memory folder across all its jobs, so the account manager agent remembers the client in week twelve.
Connect a Claude or Codex subscription from the settings page, or use API keys. Sign-in runs in the same sandbox as the agents.
Agents answer in Slack, open pull requests on GitHub and read ad and analytics data through Windsor.ai.
Read and write access to the platforms an agency already runs on, each behind the same approvals.
Ad and analytics platforms connect through Windsor.ai. Model access through Anthropic and OpenAI, on your account.