One board for every running agent
Run multiple AI agents at once and see them on one board with delivery modes, presence-aware routing, and deduped sessions.
If you run more than one AI agent, you already know the failure mode. Three Claude Code sessions in three terminals, a Codex job on the side, and your phone buzzing every few seconds with notifications you cannot tell apart. Which agent asked? What was it working on? Did you already answer that one?
The Fleet board fixes the tracking problem. It puts every running agent on a single screen, groups concurrent sessions, and routes notifications based on where you actually are.
The problem with running a fleet
One agent is easy. You start it, you watch the terminal, you answer when it asks. Notifications are barely necessary because you are sitting right there.
The trouble starts when you run several at once, or when you walk away from the desk. Now the questions stack up. A plain notification stream gives you a wall of "Agent needs your input" with no way to tell a git push approval from a database migration from a question about which file to edit. Worse, the same session can fire more than once as it retries or re-prompts, so you see what looks like five separate agents when it is really one agent asking the same thing.
You end up with two bad options. Mute everything and miss the approval that actually mattered, or leave it loud and get trained to tap through prompts without reading them. Tapping through prompts is how an agent ends up running a command you never would have allowed.
Delivery modes
Every agent connection runs in one of four delivery modes. The mode decides whether a given event reaches your phone, stays in the terminal, or both.
- Smart is the default. It uses presence to decide. If you are at the terminal, questions stay there and your phone stays quiet. If you have stepped away, they get pushed. More on this below.
- Always push sends everything to your phone regardless of where you are. Use it when you want a record on your device for every decision, or when you are running agents on a remote box you never sit in front of.
- Quiet holds notifications back. The agent still logs to the audit trail and still waits for an answer where a policy requires one, but it does not buzz you. Good for a long batch job you plan to check on your own schedule.
- Terminal keeps everything local. Nothing goes to the phone. This is the mode for when you are heads-down at the desk and do not want a second screen lighting up.
You set the mode per connection, so a production agent and a throwaway local experiment do not have to share one policy.
Presence-aware routing
Smart mode is the one most people leave on, and it depends on knowing whether you are at the terminal.
Pushary tracks terminal presence on the server, weighted by recent prompt activity. When you are actively typing prompts and answering in the terminal, the session is considered present, and Smart mode keeps questions local. When that activity goes quiet, presence decays, and the same questions start routing to your phone instead.
You never touch the toggle. Sit down and work, your phone stays silent. Get up for lunch with three agents still running, and the next approval lands on your lock screen. Presence is computed server-side from prompt weighting, so it works without shipping a new CLI build or running a background daemon on your machine.
Presence routing only changes where a notification goes. It never changes whether an action is allowed. Permission policies decide that, and they run the same way whether you are at the desk or on a beach. See permission policies.
The Fleet board
The board is where concurrent sessions become legible.
Each running agent shows up as a card in a lifecycle column, so you can see at a glance what is waiting on you versus what is still running versus what has finished. Cards carry a real task title, not a generic "session 4f2a." The title comes from the UserPromptSubmit hook, which captures what you actually asked the agent to do, so the card reads like "refactor the billing webhook" instead of a hash. Each card also shows the agent name, so a Claude Code session and a Codex session are never confused.
Concurrent sessions are grouped and deduped. If one agent re-prompts or fires the same question twice while it retries, the board folds those into a single entry rather than showing you duplicates. That is the difference between a board that says "one agent is blocked" and a stream that makes it look like your whole fleet caught fire.
So the board answers what a notification stream cannot: which agent, what task, what stage. From there you approve, deny, or send a correction back to the agent, all from the same view.
How to start
The Fleet board is part of the AI agent control panel and works with any agent you connect through the CLI hook: Claude Code, Codex, Gemini CLI, Cursor, and Hermes. Connect an agent with the quickstart, run a couple of sessions, and they show up on the board with titles and names already filled in. Set the delivery mode per connection from the Connections settings tab.
One honest caveat on mobile. On an iOS home-screen install, the cross-origin deep link into the app is broken, so iOS surfaces pending questions through an inbox on the subscribe page rather than a tap-through link. Android and the native app handle the deep link directly.
The board, presence routing, and delivery modes come with every paid agent plan. See the agents overview for the full feature set and pricing for the two tiers. Agent is $9.99 a month, Agent Pro is $19.99.