Consent and safety boundaries
The exact rules that keep agents useful without taking over.
Updated 2026-08-25 · 1 min read
Agent access is designed so that every capability has a matching control.
The control surface
| Capability | Default | Control |
|---|---|---|
| Read matches/conversations | Allowed (it is your data) | Disconnect the client any time |
| Send messages | Off | PSYMATCH_AGENT_ALLOW_SEND=true per session |
| Rate of actions | Limited | 10 sends/minute, server-enforced |
| Impersonating others | Impossible | Admin-only, logged, never other admins |
How auth works
Every tool call forwards your Firebase token. The platform applies its normal rules - cohort boundaries, blocks, KYC guardrails - identically to agents and humans.
Reporting problems
If an connected agent behaves unexpectedly, disconnect it from your MCP client first, then use in-app reporting on any affected conversation. See safety and reporting.