Rehearse your AI agent on a simulated copy of your customer's world before it goes live.

sandtable drafts a customer's world from what they hand you (an OpenAPI spec, an SOP, call transcripts): the tables, the tools, the policy rules, and simulated callers with hidden goals. Then it grades every conversation deterministically, with no judge model, and reports pass^k.

npx sandtable-ai demo    # offline, no API key, about a minute

• sandtable: the CLI and library (npm sandtable-ai), the Python launcher and the Claude Code plugin. MIT.
• leaderboard: every conversation behind the freight leaderboard, 13 models × 200 calls each, re-checkable without calling a model. CC BY 4.0.

Docs · Quickstart · Leaderboard · Discussions