The trust layer that runs under everything

Supervision

We don’t fail silently — most agents do, and you find out after the deal is lost. Supervision is the moat: a configurable human checkpoint before every external action, and a write-only record of all of it.

Mechanism 01 · autonomy dial

Supervised by default

A per-stage dial governs send, reply and book — each defaulting to a human approval checkpoint. You open autonomy stage by stage, exactly where you trust it. Try it below.

Mechanism 02 · audit log

On the record, always

A write-only audit log records one row per mutating action, indexed by tenant, customer and action — framed against DPDP §8. Nothing the operator does externally goes unlogged.

write-only · append-only
indexed: tenant · customer · action
reply.sent · book.requested · lead.qualified
framed against DPDP §8

try it — the autonomy dial

Supervised by default.

Drag the handle to open autonomy where you trust it — send, then reply, then book.

sendthe outbound message

human-approve

replythe next reply in a thread

human-approve

booka calendar booking request

human-approve

dial.audit_log

— no changes yet · drag or arrow the dial to configure

local sandbox — nothing here is sent

04 — everything else under the hoodCircuit-breaker LLM fallback chainDurable inbound · re-enqueued on crashInput / output / model guardsConsent gateWABA guardTiers & meteringMCP + A2A endpoints~787 tests passing

supervision you can show your customers

See it run on your own front office

Book a walkthrough → (opens your email app)Back to home

30 minutes, free — we run it live on your own front office, not a slide deck · services@thrvin.com