AI agents you can
hand the keys to.
Treza is the runtime for autonomous agents that handle money, data, and real-world decisions. Hardware-isolated execution, a tamper-evident log of every action, and key custody your agent can use but never leak.
- 0 keysEver leave the enclaveYour agent signs, the secret never does
- 100%Of agent actions are loggedTamper-evident, byte-level, replayable
- Any modelOpenAI, Anthropic, open-weightCall inside or outside the enclave
- ~2 minFrom container to running agentBring a Docker image, get a runtime
Built for agents that do real work
Pick the agent pattern closest to yours. Each one ships with the primitives, integrations, and reference code you’d expect.
Data & Healthcare Agents
Agents that read PHI, PII, or proprietary datasets and call models on them without exposing the underlying data.
- AI over PII / PHI
- HIPAA-ready patterns
- Model never sees raw data
Identity & KYC Agents
Agents that verify users, screen for AML risk, and gate access without storing the identity data they checked.
- ZK-KYC primitives
- Pass-through verification
- Auditable approvals
Treasury & Payment Agents
Agents that move money, manage stablecoin treasuries, or settle invoices on behalf of a finance team.
- Sealed signing keys
- Spend policies in-enclave
- Per-action attestation
Wallet & Custody Agents
Agents that approve, sign, and settle on-chain. Their keys live in the enclave; their decisions live in an auditable log.
- Sealed signing keys
- Approval policies in-enclave
- Replayable on-chain actions
From idea to trustworthy agent in 3 steps
No enclave expertise. Bring the agent stack you already use. Treza handles the runtime, key custody, and proof.
- Step 01
Define your agent
Bring any agent stack you already use. Containerise the code, declare the tools, models, and keys it needs. No SDK rewrites.
$ treza agent init my-treasury-agent - Step 02
Treza runs it in a sealed enclave
Your agent boots inside a hardware-isolated runtime. Its keys are generated in-enclave and never extractable. Even your cloud operator can’t peek in.
$ // Online. Sealed. Attested. - Step 03
Prove every action it took
Every tool call, every signature, every model invocation is logged with a cryptographic receipt. Replay any run, audit any decision.
$ GET /v1/agents/{id}/actions
Why autonomous agents need a different runtime
A short animated explainer of why agents that handle money and data can’t live on a regular cloud, and how Treza fixes it. Want a walkthrough of the runtime itself? Book a demo.
Everything your agent needs to be trusted
Six runtime primitives. One platform.
Compose them into agents your CFO, GC, and CISO will all sign off on.
Sealed key custody
Your agent generates and uses its own keys inside the enclave. It can sign anything you authorise, and extract nothing. Ever.
Learn more->Tamper-evident action log
Every tool call, every model invocation, every signature is recorded with a cryptographic receipt. Replay any run, prove any decision.
Learn more->Private model calls
Your agent can call OpenAI, Anthropic, or open-weight models on PII or PHI. The provider sees a request shape, never the underlying data.
Learn more->Policy-bound execution
Declare what your agent is allowed to do (spend limits, allowed counterparties, hours, kill-switches) and have the runtime enforce it in-enclave.
Learn more->Identity-aware tools
Drop-in ZK-KYC primitives so your agent can verify users, screen for AML risk, and gate access without ever storing the underlying ID.
Learn more->Any model, any chain, any tool
Bring the LLM, framework, and tools your team already uses. Treza wraps them in a runtime your security review will actually approve.
Learn more->The runtime, as a platform
The Agent Runtime is the first product built on Treza Platform. The same primitives are available to teams who want to build their own runtime, or wrap a vertical-specific one.
Drops into the stack you already run
Ship containers you already build, sign for the chains you already use, and call the models you already trust, without changing how your team works.
Agents we’ve
shipped with teams
Real agent patterns running on Treza today, from treasury and trading to identity verification and clinical workflows.
Signing agents
Agents that hold their own signing keys, enforce spend policies in-enclave, and emit a cryptographic receipt for every transaction they authorise.
Trading agents
MEV bots and execution agents whose strategy stays hidden from the operator and whose signatures can’t be exfiltrated.
Clinical agents
Read PHI, call any model, return decisions. The model never sees the raw record.
Audit-ready agents
Every agent action is signed by the enclave. Replay any decision, export any trail.
Data agents
Agents that process customer data, contracts, or media without ever exposing the raw bytes upstream.
KYC & identity agents
Agents that verify users, screen for AML risk, and gate access to your product. Treza ships ZK-KYC primitives so they can prove a check passed without storing the underlying ID.
Common questions
What teams ask before they ship their first agent on Treza.
What is an agent runtime?
It’s where your agent actually executes. Treza gives your agent a hardware-isolated sandbox with its own keys, its own memory, and a tamper-evident log of every action it takes. Think of it as the gap between “an LLM with tools” and “an agent you can let touch production.”
Does Treza replace LangChain, Mastra, or my own framework?
No. Treza is the runtime underneath. Bring whatever framework, model, or tool stack you already use. Treza wraps it so your agent gets sealed keys, a sandboxed environment, and an auditable trail without you rewriting anything.
Can my agent hold and use private keys?
Yes. Keys are generated inside the enclave, used to sign whatever you authorise (transactions, payloads, attestations), and never leave. Even your cloud operator can’t extract them.
How do I prove what my agent did?
Every tool call, model call, and signature is logged with a cryptographic receipt anchored in the enclave attestation. You can replay any run, export any action, and hand a regulator hardware-signed evidence instead of a screenshot.
Which models can my agent call?
Any of them. OpenAI, Anthropic, open-weight models, or your own fine-tunes. Inference can happen inside the enclave (model never sees raw data) or via a proxied external API.
Which clouds does Treza support?
Treza runs on the major confidential-compute clouds and is designed so the same agent definition works across providers. We can map your specific requirements during a 20-minute architecture review.
Bring your agent. We’ll design the runtime.
Tell us what your agent does (or should do). In 20 minutes we’ll sketch the runtime, the keys it needs, and what production looks like. No pitch deck.