Esclusa

A safety gate for AI agents that change infrastructure.

checking chain…

A coding agent asks the gate before it touches your infrastructure. The gate answers against the live service topology, signs the decision into a ledger that cannot be quietly edited, then keeps probing the network to catch whatever happened anyway.

  1. 1 Evaluate The agent proposes an action. The gate allows or refuses it, and records either way.
  2. 2 Observe Probes run every 15 s from inside the private network, catching changes that skipped the gate.
  3. 3 Prove Every entry is HMAC-chained, so editing the record breaks it at that row.

Try it

three things a coding agent might do to this project, right now
1

Delete the database

An agent decides db is unused and tries to drop it.

expect refused instant
2

Retire a service, properly

Ask the gate first, then really take oldworker offline.

expect audited ~45 s
3

Go behind the gate's back

Take oldworker offline with no approval on record.

expect drift ~45 s

Scenarios 2 and 3 genuinely stop oldworker answering for 90 s — the probes below are failing for real. It restarts itself, and it exists to be knocked over.

Topology

live probes inside the private network · every 15 s
awaiting gate…

Ledger

tamper-evident decision chain · refresh 5 s
timekindactoraction decisionreasonpolicyhmac
awaiting gate…