Prove what your AI actually did
Juratify records every important action your AI agents take and turns it into tamper-proof evidence, so you can meet EU AI Act record-keeping duties and show an auditor exactly what happened, months later, without taking anyone's word for it.
Under the hood: cryptographically chained records, sealed on your side with a key we never see, verifiable offline by anyone. See the architecture.
Works with your existing AI stack
Two ways in, and neither asks you to change how your agents are built.
OpenTelemetry
Already tracing your agents? Point your OTLP exporter at Juratify, or drop in our span exporter. We read the GenAI semantic conventions, so chat calls, tool calls, and errors arrive as evidence without rewriting your agent.
The Juratify SDK
One call records prompts, completions, tool calls, and token usage from any TypeScript agent. Payloads are encrypted in your process before anything leaves it, so the vault stores ciphertext it cannot read.
Agents built on these reach the vault through OpenTelemetry traces or the SDK. Model and framework names are listed to show what we ingest from, and do not imply any partnership or endorsement.
How it works
The SDK seals every event
Your SDK hash-chains each record and encrypts the payload with a key we never see, then ships only ciphertext to the vault.
The vault signs receipts and checkpoints
On ingest the vault signs a receipt for every event and periodic Merkle checkpoints over the whole chain, anchoring your records against later tampering.
Anyone verifies it offline
Export a signed evidence pack and re-verify it with the open-source verifier. No account, no network, and no need to trust us.
The right events, without slowing your agent
A logging tool is only worth as much as what it captures and what it costs you to capture it.
The events that matter in an audit
Session starts and ends, prompts, completions, tool calls, guardrail triggers, errors, escalations, and your own custom events. Each one is filed automatically against the Article 12(2) record-keeping category it evidences, so the export reads as a compliance record rather than a log dump.
Off your agent's critical path
Sealing and delivery happen outside the call your agent is waiting on. A write-ahead log on disk absorbs crashes, restarts, and our own downtime, then drains when the vault is reachable again, so recording never blocks a model call and never silently drops one.
Private by construction
In client-sealed mode the payload is encrypted with a key held in your environment and never sent to us, so sensitive prompts stay unreadable to Juratify while their integrity stays provable to anyone. Model ids can be redacted per workspace.
Why you can rely on it
Built for teams shipping AI on OpenAI, Anthropic, and Azure. Every claim below is one you can check yourself, today.
Open where it counts
The verifier and the SDK are Apache-2.0 on GitHub. An auditor runs npx @getjuratify/verify against your evidence pack with no Juratify account and no network access, so the proof does not depend on trusting us or on us still being here.
We cannot read client-sealed payloads
The data key stays in your environment. What we hold is ciphertext plus the chain metadata needed to prove nothing was altered. That is a property of the design, not a promise in a policy document.
Append-only, signed, and anchored
Events are write-once at the database level. Every event gets a signed receipt on ingest, and periodic Merkle checkpoints anchor the whole chain, so edits, deletions, and reordering are all detectable after the fact.
Try it before you commit
Every plan starts with a 14-day free trial, and paid charges carry a 14-day refund window. Export your evidence at any time in an open, documented format that verifies without us.
Juratify does not ask you to take its word for anything: the format is open, the verifier is open source, and your evidence verifies without us. Read the source.
Frequently asked questions
What is Juratify?
Juratify is evidence-grade, audit-ready logging for AI agents. Every event is hash-chained, signed with a receipt, and anchored by periodic Merkle checkpoints, so the record is tamper-evident and anyone can verify it offline. It is designed to support EU AI Act Article 12 and Article 26 record-keeping.
What does "evidence-grade" mean here?
It is a technical integrity claim, and a specific one: every event is hash-chained to the one before it, individually signed with a receipt on ingest, anchored by signed Merkle checkpoints, and re-verifiable offline by anyone with an open-source tool. Tampering is therefore detectable, and the integrity of the record can be demonstrated to a third party rather than asserted. It is not a certification, and it is not legal advice. Whether a record is admissible in a given proceeding also depends on your own procedures and your jurisdiction; Juratify gives you the technical foundation to make a chain-of-custody argument, not a guarantee about the outcome.
What does the EU AI Act require me to log?
For high-risk AI systems, Article 12 calls for automatic recording of events over the system's lifetime, and Article 26(6) asks deployers to keep those automatically generated logs for at least six months, subject to applicable law. Juratify records the events, keeps them tamper-evident, and retains them for at least that long. It is a tool to support those obligations, not legal advice, and it does not by itself guarantee compliance.
Can Juratify read my logs?
For client-sealed events, no. Payloads are encrypted in your own environment with a data key you hold and never transmit to us, so we store ciphertext we cannot read. This is a structural property of the design, not a policy promise.
How does an auditor verify the evidence?
Export a signed evidence pack and run the open-source verifier offline with npx @getjuratify/verify. It rebuilds the hash chain and checks every receipt and checkpoint signature against the published keys, with no Juratify account and no network access.
What happens if someone tampers with a record?
Editing, deleting, or reordering any event breaks the hash chain, and the signed checkpoints anchor the chain so later changes are detectable. Verification then fails and points to the exact event where the chain broke.
What does it cost, and is there a trial?
Plans start at EUR 99 per month and scale by workspace count, event volume, and retention. Every plan begins with a 14-day free trial, and paid charges carry a 14-day refund window. See the pricing page for the full breakdown.
What is client-sealed versus server-sealed?
Client-sealed events are encrypted in your environment with a key we never see, so only you can decrypt the payloads while the integrity metadata stays verifiable by anyone. Server-sealed events arrive through the OTLP door and are encrypted at rest with a per-workspace key we hold, which helps when you cannot seal on your side. You choose per integration.
Read more about our security architecture, or see the pricing.