Krynix · Independent evidence for AI agents
Prove your agents stayed in bounds.
Krynix is the independent, tamper-evident record of every decision your AI agents make — held by a neutral third party, verifiable by anyone. Bring your own enforcement (Microsoft AGT, LangChain, our gateway); Krynix holds the evidence your own logs can't prove.
The problem
Your agents act now. Can you prove they stayed in bounds?
Blocking bad actions is becoming a commodity — even Microsoft ships it free. The unsolved problem is proof: evidence an auditor, a regulator, or your customer’s security team will actually trust — not a log the operator can edit.
“Each agent action looks legitimate on its own — together they quietly exfiltrate data.”
See deterministic enforcementCan you prove it?“An auditor wants tamper-evident proof our agents are governed — our own logs are just our word.”
See the evidence layerAcross your stack“We already block bad actions — but we can't independently prove what was stopped, or why.”
See how it fitsHow it works
Emit. Hold. Attest. Verify.
Your agents — and whatever you use to govern them — emit decisions. Krynix holds them in an independent, tamper-evident record and turns them into evidence anyone can verify. We sit above your stack; we don’t replace it.
Emit
Whatever you already use to govern agents — Microsoft AGT, LangChain, your own code, or the Krynix gateway — sends its decisions to Krynix over OpenTelemetry or a one-line HTTP contract.
Hold
Every decision lands in an independent, append-only store and is hash-chained on arrival, so any later edit or deletion is detectable. One immutable record per agent, written once.
Attest
Records are hash-chained and normalized into one evidence model across every agent stack you run — not one silo per framework. Cryptographic signing and external anchoring are on the roadmap.
Verify
Turn the record into auditor-ready evidence packs, mapped toward the formats regulators expect — ISO 42001, the EU AI Act audit file, FINRA records retention.
Design principles
Built on three principles that decide what Krynix does and what it deliberately won't.
Independent custody, not self-attestation
A log the audited party generates and holds is just their word. Krynix records are held by a neutral third party — the operator can read and query them, but cannot alter or delete history. Separation of duties is the entire design, because it is what makes the record count as evidence to someone outside the company.
Sit above your stack, not replace it
Your existing enforcer blocks the action; Krynix holds the proof. Decisions arrive over open standards and are normalized into a single evidence model across every framework you run — so the control plane stays yours, and the evidence stays neutral.
Verify it yourself, don't trust us
Records are tamper-evident and checkable with an open verifier, against an open format — trust is structural, not a promise. And by design we capture structured, non-sensitive evidence only: never raw prompts, argument values, or PII, so it is easier to retain and to defend.
Get started
From your existing stack to independent evidence in four steps.
Keep your enforcement
Already block bad actions with your own enforcer, a framework's guardrails, or our drop-in gateway? Keep it. Krynix doesn't replace your control plane — it sits above whatever you run.
Point decisions at Krynix
Send your agents' decisions to Krynix over OpenTelemetry or a one-line HTTP contract. No code rewrite, no per-framework wiring, no lock-in. If you have nothing yet, the Krynix gateway is a transparent MCP proxy that emits for you.
Get independent evidence
Every decision lands in a tamper-evident, independently-held record — the operator can query it but can't alter history. Normalized into one evidence model across every agent stack you run, with the causal 'why' behind each block.
Prove it to anyone
Export auditor-ready evidence; anyone can verify a record's integrity with the open verifier, no trust in us required. Become a design partner and we'll shape the evidence format to exactly what your regulator or customer asks for.
Early access
Become a design partner.
Two minutes, seven questions. We’re early and selecting a small group of design partners — companies deploying agents that touch sensitive data or take real actions. We reply within five business days.
- Shape the evidence format around your auditor's requirements.
- Priority Slack / Discord channel with the team.
- Hands-on help wiring your first integration.
Powered by Tally · Responses → [email protected]
Or just say hi
To reach us.
Not ready for the form yet? Drop into the community, follow along, or send us a note about what you are trying to govern.