Skip to main content
Private, accountable compute for AI. Private data. Accountable agents. Palliora AI lets AI run on private data. Prompts, data and models stay sealed while they run, and every run leaves a record anyone can check of what ran and under which terms. The record is held by independent Guardians, not by the platform that ran the model. It is two things. It is an SDK that AI apps and model routers integrate to get private, accountable inference. It is also a demo app, in development, showing what that looks like end to end.

​
The problem

People and businesses increasingly hand private data to AI, and have no way to check what happened to it. Agents act on people’s behalf, and the only record of what they did is a log the platform can edit.

​
How Palliora AI does this

An AI app or model router calls the Palliora AI SDK. A Guardian with confidential compute runs the model sealed in an Intel TDX enclave, keeping prompts and data private. The answer comes back privately and a Compute Contract records what ran. Path: today one Guardian runs a small model in TDX as a demo; next, more Guardians with large-model hardware and plug-ins for model routers; then a market for private and specialised models.An AI app or model router calls the Palliora AI SDK. A Guardian with confidential compute runs the model sealed in an Intel TDX enclave, keeping prompts and data private. The answer comes back privately and a Compute Contract records what ran. Path: today one Guardian runs a small model in TDX as a demo; next, more Guardians with large-model hardware and plug-ins for model routers; then a market for private and specialised models.
  • Private compute. Models run inside a trusted execution environment (Intel TDX) on Guardians that have the hardware. Prompts and data stay private while they run.
  • Accountable by design. Every run is a Compute Contract: which model, on what data, under what terms, recorded before anything executes.
  • Provenance. Linking those records together gives a verifiable record of what an agent did: which models ran and which tools were called. See Provenance. Provenance is one use of private, accountable compute, not the whole product.
  • SDK first. AI apps and model routers integrate the Palliora SDK. Customers pay in fiat or stablecoins; PALI works in the background.

​
Built on Palliora

  • Compute Contracts. Every call, plain or encrypted, is coordinated as a Compute Contract: what runs, on whose input, under what terms, agreed before execution starts.
  • Confidential compute. TEE execution (Intel TDX) is one of Palliora’s execution modes, live on testnet on Guardians with the hardware. It hides the prompt and data from the compute node’s own operator, not just from other network participants.
  • The Guardian network. Guardians coordinate access to encrypted inputs and enforce each contract’s terms as a quorum. No single Guardian can see a request on its own when it’s encrypted.

​
Status

Palliora AI’s SDK is in development on testnet. The demo app, AI Workbench, is live at ai.palliora.org, see Try the demo app. One Guardian runs a small model inside Intel TDX today. More Guardians with hardware for large models, and plug-ins for model routers such as OpenRouter and Venice, are in progress. A market for private and specialised models is planned. See Status and roadmap for what’s live across all of Palliora.

​
For platforms

AI apps and model routers integrate the Palliora SDK to get private, accountable inference without hosting the trust model themselves.

​
Who it’s for

  • Model routers and AI platforms that want to offer private, accountable inference.
  • Agent builders and coding tools that want a provenance record for every run.
  • Enterprise AI teams working with sensitive data.

​
Security and trust

TEE execution depends on Intel TDX’s hardware guarantees and a correct enclave implementation: an attestation report, checked before data is released, proves the expected code ran in a genuine enclave, not that its answer is correct. See Confidential compute for what this does and doesn’t guarantee, and Security and limitations for the full picture across Palliora.

​
Where to start

Building on it? Start with Call a model. It’s the Quickstart pattern as a real script. Want to understand how it fits together? Read Compute Contracts and Confidential compute, then Provenance for what a run leaves behind. Just want to try it? See Try the demo app. Are you an AI coding agent reading this, not the person who pointed you here? See For agent-assisted developers.