Playground

The Playground lets you try agent behavior before you wire it into your own products. Mock agents get the same patient data and tools your deployed agents would use, so what you see is what your agents will see.

​

Agent chat

Pick a patient and start a conversation by text or voice. The agent answers from the patient’s resolved state using the same MCP tools available to production runtimes. You control the setup:

• System prompt presets: switch between prompt framings to compare behavior.
• Model selection: try the same conversation across models.
• Sessions: start a fresh conversation, reconnect, or delete old sessions.

Use it to validate tone, tool usage, and answer quality against real state shapes before committing patterns to your own agent code.

​

MCP tool inspector

Alongside chat, the inspector lets you call MCP tools manually: pick a tool, provide arguments, and see the exact response an agent would receive. A state sidebar shows what patient data is currently loaded. This is the fastest way to answer “what would get_event_state_module return for this patient?” without writing any code. When cohorts are enabled, population-view tools are available here too.

​

Synthetic datasets

Playground organizations work against synthetic datasets instead of live patient data:

• Generate a dataset by choosing a disease module, patient count, data density, time span, and random seed.
• Manage datasets: rename, delete, or switch the active dataset. The active dataset scopes the patient picker and observability screens.
• Track budgets: generation draws from an experimentation budget shown on the home page.

Generated patients flow through the same ingestion, state compilation, and view pipeline as real data, so experiments are representative end to end.