Open runtime governance for AI agents

Let agents act. Decide what they're allowed to do, the moment they do it.

Deadlatch is an open stack of three primitives. Purse enforces what an agent can do, blackbox proves what it did, Tripwire watches for what slipped through. Install one from npm, or run all three.

You can't trust a black box to govern your black box.Every part is open, inspectable, and verifiable outside Deadlatch. No platform to take on faith.
enforce
Purse
prove
blackbox
watch
Tripwire
DEADLATCHlive consoleenforcing
max/action $100max/day $200approval > $50allow api.stripe.com, *.aws
agent → spend $12.00 to api.stripe.com
allowwithin policy
daily spend$0.00 / $200.00
blackbox · hash chainprove
verify() ✓ ok
tripwire · watchdetect
0actions caught
Three primitives, one control loop

Prevent the wrong action. Prove what happened. Detect what slipped through.

Each is a small, open package you can adopt on its own. Together they close the loop around every action an agent takes.

enforce · Purse ↗

The action never fires off‑policy

the credential lives where the agent can't reach it

Route every spend or tool call through Purse. It checks the action against live policy at the moment it happens, and a hijacked agent still cannot move money outside the rules.

const d = purse.authorize(({ amount: "$80.00" }) // d.status → "needs_approval"
prove · blackbox ↗

A record no one can quietly edit

hash‑chained, verifiable outside the tool

Every decision is written to a tamper‑evident log. Edit, insert, or reorder a single record and the chain breaks. verify() names the exact record that was touched.

box.append(({ action, verdict }) box.verify() // → { ok: false, brokenAt: 4 }
watch · Tripwire ↗

Catch the silent wrong turn

read‑only, before a customer does

Tripwire watches an agent run and flags the action that shouldn't have happened. It changes nothing, so you can put it beside a live system today.

tripwire.watch(agentRun) // flags the off‑policy action → alert
You can't trust a black box to govern your black box.
01

Open, not a platform you take on faith

Every primitive is source you can read and run. The thing enforcing your policy is not itself a mystery box.

02

Verifiable outside the tool

The audit chain checks out with plain SHA‑256, on your machine, without Deadlatch in the loop. Proof you hold, not proof we assert.

03

Composable, adopt one at a time

Start with the one primitive you need this week. Grow into the full loop when you're ready. No rip‑and‑replace.

Why now

The rest of the field is arriving at the same three controls.

This is not our claim. Independent security guidance, a US Senate draft, and this year's incident data all point at one loop. Enforce what an agent can do, prove what it did, watch for what slipped through. The sources are named so you can check them.

SANS · practitioner ↗

"Your AI agent is an easily confused deputy. Cloud security needs a credential broker."

enforce · a credential broker is Purse
AI AGENT Act · US Senate draft

Calls for scope-limited delegation credentials, real-time revocation, and auditable records.

enforce + prove · grants, revocation, receipts
HAID framework ↗

Proposes binding an agent's permissions to human-declared intent, not a coarse scope.

intent-binding · Purse binds the exact action
CSA + vendor reporting · 2026

A majority of organizations running AI agents reported an agent-caused security incident this year.

why now · the failures are already happening

Honest about the receipts. The AI AGENT Act is a discussion draft, not law, and the incident figures are vendor-reported. We link and name them so you can weigh them yourself, the same standard the audit chain below holds itself to.

Evidence you can hand an auditor

The controls the new AI rules ask for. As proof you can verify yourself.

Human oversight, record-keeping, and monitoring. Deadlatch gives you the actual controls, open, and evidence an auditor can check without trusting us.

Human oversight of risky actions
EU AI Act · Article 14
Purse
enforce
Automatic, tamper-evident record-keeping
EU AI Act · Article 12
blackbox
prove
Ongoing monitoring for unsafe behaviour
NIST AI RMF · Manage
Tripwire
watch
No compliance checkbox to take on faith. You get the open controls and evidence your auditor can verify outside the tool.Aligned with the record-keeping, human-oversight, and monitoring expectations of frameworks like the EU AI Act and the NIST AI Risk Management Framework. Deadlatch provides controls and evidence, not a certification.
Start with one

One npm install. No account, no platform.

Pick the primitive that solves today's problem. Each runs on its own, zero dependencies.

Purse
enforce
npm i @olurabian/purse
blackbox
prove
npm i @olurabian/blackbox
Tripwire
watch
npm i @olurabian/tripwire
Then wire the action through it. Prevent. Bear witness. Track.
fail closed
Closed at the moment of action.