Skip to main content
Pome is simulation testing infrastructure for AI agents. Stateful digital twins answer the same REST and MCP calls as production systems, so you can evaluate agent behavior end to end without touching real users, data, or APIs. A scored run needs three inputs: a seeded world state, a task prompt, and the twins the agent can call. Vary those, and Pome surfaces the edge cases unit tests miss. Live runs and scores land on app.pome.sh.

​
Quickstarts

Log your first scored run

Paste one prompt into Claude Code and get a failing score you can fix.

Connect your agent

Point Pome at your repo — one paste-prompt writes pome.json and registers it.

Drive it from MCP

Author tasks, launch runs, and read reports by tool call from your coach.

​
Workflow

How Pome works

What happens in a run, from seed state to scored report.

Which surface?

CLI, skill, MCP, dashboard — same loop, pick where you drive it.

Dashboard

Agents, twin sessions, traces, and judge handoffs at app.pome.sh.