Constitutional Core
Five inviolable axioms, SHA-256 sealed at boot. All-or-nothing: if one fails, the task is rejected — no override, no profile, no exception.
The seatbelt for autonomous AI
Safety infrastructure for autonomous AI.
Kovrin sits between your AI agents and the real world — enforcing constitutional limits, routing risk deterministically, pausing for human approval, and recording every decision in a tamper-evident audit trail. It wraps LangGraph, CrewAI, or any agent stack.
Drops into LangGraph · CrewAI · your own agents
What Kovrin is
Kovrin is not a chatbot framework, and not an all-in-one runtime. It's safety infrastructure — like a seatbelt, not the car. Your agents keep their autonomy; Kovrin makes sure they can't cross a line they shouldn't.
- 5
- Constitutional axioms, SHA-256 sealed
- 8
- TLA+ specs · 10 proven invariants
- 4×3
- Deterministic risk-routing matrix
- Append-only
- Merkle audit, tamper-evident
How it works
One pipeline. Every intent goes through it.
No side doors. Every request — from a human or an agent — is parsed, judged by independent critics, checked against the constitution, routed by risk, executed, and sealed into the audit chain.
- 01 Intent
- 02 Critics
- 03 Constitutional Core
- 04 Risk Router
- 05 Execution
- 06 Merkle Audit
Why it's differentLayer 0 
Routing
Audit
Proven
Execute
Control
Guardrails that are built in, not bolted on.
Deterministic risk routing
Every action is scored and routed by a fixed matrix. CRITICAL risk always routes to a human — hardcoded, not configurable.
Tamper-evident trail
An append-only Merkle hash chain. Every decision traces back to the original intent — and any tampering is detectable.
Formal verification
Eight TLA+ specifications, ten invariants. The safety properties are mathematically proven, not just tested.
Orchestration with guardrails
Plans decompose into a DAG and run in parallel — always inside the safety envelope, never around it.
Human-in-the-loop by default
Irreversible or high-risk actions pause for explicit approval. Autonomy is opt-in per capability — never assumed.
Designed for the regulated world.
Aligned with EU AI Act · Articles 9 · 12 · 14 · 15
Get in touch
Want Kovrin guarding your agents?
Tell us about your agent stack and what you need to keep under control. We'll show you how Kovrin plugs in.
hello@kovrin.dev