Demo
A real LangGraph agent, backed by a real model, making real writes against a live Postgres through nothing but Terminus. No narration, just what actually happened.
What you are watching: the agent's only database access is the Terminus MCP
server's query and
execute tools. An allowed
write executes, a destructive write is denied and the row survives, a wildcard read is denied
and the agent self-corrects using the rewrite Terminus hands back, and the session's
tamper-evident audit chain verifies, then breaks the moment one decision is altered. Every
verdict is real: it comes from actual database state, not the agent's account of itself.