MadData Labs
Scenario
Clean monthly close
Conversation
Ready
No turns yet

Pick a starter prompt below or write your own. The agent will pull from the GL and subledger MCP servers and stream its reasoning back here.

Starter prompt
Run a full month-end reconciliation for the Apr 30 close across all asset classes. Produce an exception report.
Artifacts
Reconciliation breaks will appear here once the agent finishes its first comparison pass.
What’s happening

About this scenario

GL and subledger tie out across all asset classes. Walk through the agent's plan, the tool calls it makes, and the structure of its final reconciliation report.

What Anthropic runs

Anthropic
  • The agent loop, in a cloud-hosted container.
  • Persistent session state and the event stream.
  • Tool execution for built-ins (read, write, edit, glob, grep).
  • Prompt caching and automatic compaction.

What we run

MadData
  • The MCP servers (GL and subledger), in this Next.js app.
  • The system prompt and trust-tier policy.
  • The synthetic data and the scenario mutations.
  • This UI — the bridge that turns Anthropic events into AI SDK message parts.

Trust tiers in action

Watch the tool calls. GL and subledger MCP responses are trusted. Any document body the agent reads via sub_fetch_statement is untrusted — its content is data, never instructions.