MadoAgent

MadoAgent is the conversational side of MadoHub. Open it with Cmd+Shift+M (⌘⇧M) and describe what you want done — "set up a peer review between two Claude agents on this API", "try two caching strategies on this bug and tell me which one wins" — instead of building it tile by tile yourself.

It is not a chat wrapper around a single model call. Every message runs a real tool-calling loop that can see your canvas, spawn and message agents, wire routing connections, open frames and terminals, and read back its own memory. In-app onboarding calls it "your orchestration brain," and that is the accurate framing: MadoAgent is a peer to the canvas, not a suggestion box.

What you can do with it

• Spawn and assign agents — start Claude, Codex, or Amp in a fresh terminal, or hand work to an idle agent of the right type in the same project instead of stacking a duplicate.
• Wire agents together — ask for a peer review, pipeline, or watcher pattern and MadoAgent draws the routing connection with the right trigger and loop settings.
• Drive the canvas — open frames and terminals, create notes, jump to a specific tile, or clear the whole canvas from a single prompt.
• Set up environments — open a project and run a sequence of setup commands (install, start dev server) for you.
• Run an autonomous terminal — point MadoAgent at a terminal and a goal, and it will keep issuing commands and watching output until the goal is done.
• Explore alternatives in parallel — spawn several agents on the same problem with different instructions and compare their results side by side (see Ensemble Exploration).

How to use it

1. Open the panel with Cmd+Shift+M (⌘⇧M).
2. Type your request in natural language and press Enter.
3. Watch MadoAgent spawn tiles, draw connections, and message agents on the canvas. It surfaces toasts and inline notes when something needs your attention (for example, a connection it refused because a named tile was not actually on the canvas).
4. Keep the conversation going — MadoAgent remembers the last 20 messages as context, so you can follow up with "now connect those two for peer review" without re-explaining.

Choose how much MadoAgent is allowed to do directly in Settings → Agent:

Before MadoAgent responds at all, you need a provider and API key configured in Settings — it shares its API configuration with agent routing.

Common use cases

• "Start a Claude agent in my project and a Codex reviewer, then connect them for peer review." MadoAgent spawns both tiles and wires the connection in one turn.
• "Dispatch a Codex worker to refactor the auth module while I keep working here." A headless worker tile gets a Worker badge so you can tell it apart from an interactive one. See Orchestrator Mode.
• "What is Claude doing in the API project right now?" MadoAgent reads a fresh snapshot of every terminal, frame, and connection before answering — no tile switching required.
• "Try two different caching strategies for this bug and tell me which wins." MadoAgent forks several agents on the same problem and lists them in the Exploration panel for you to compare.

Tips and best practices

• Prefer MadoAgent for anything that does not fit a built-in Flow — for repeatable two-agent patterns, Flows are faster.
• If a request names something that is not on the canvas, MadoAgent refuses the unsafe part and tells you in the chat rather than leaving a broken connection behind. Read the toast and rephrase.
• Use the Orchestrate skill level when you want MadoAgent to coordinate agents but keep its hands off files.
• Long-running asks like "set up the dev environment and start the server" work well with the autonomous terminal mode — check back on the tile instead of babysitting the panel.

Related features

• Flows — one-click reusable agent pairings for common patterns.
• Agent Routing — what triggers and transforms actually do; MadoAgent uses the same routing engine.
• Connections — wire agents by hand for one-off, precise control.
• Memory — long-lived facts and preferences MadoAgent can recall across sessions.

Limitations

• MadoAgent keeps only the last 20 messages as context. For very long orchestration tasks, summarize decisions into Memory so they survive.
• It depends on a working provider and API key. A failed API call retries twice with backoff before surfacing an error.