Illustration for /ap3

Introducing AP3

The Privacy Layer for the Agentic Economy

Open protocol for privacy-preserving multi-agent collaboration.
Agents across organizations, jurisdictions, and vendors compute jointly.

DESIGN PARTNERS

Decentralized trust wordmark graphic

AP3 is open sourced under Linux Foundation Decentralized Trust. View repo →

The next bottleneck is not capability.
It is trust.

Illustration for /ap3
Illustration for /ap3

Collaboration stops at trust boundaries

Individual agents are powerful, but collaboration becomes difficult when they must work across organizational boundaries.

Illustration for /ap3
Illustration for /ap3

Today’s choices are all compromises

Organizations must either share sensitive data and lose confidentiality, withhold data and lose valuable insight, or rely on a third party to bridge the gap.

Illustration for /ap3
Illustration for /ap3

Regulated decisions require a new model

These tradeoffs are unacceptable in financial services, healthcare, legal, and other domains where organizations need to reach decisions without exposing confidential information.

Three vertical panels showing cryptographic assets
Central node icon within concentric circles
Diagram of connected chat blocks

Collaborative computation with cryptographic confidentiality

Collaborative computation with cryptographic confidentiality

AP3 provides a privacy layer for secure computation across organizational boundaries.

AP3 provides a privacy layer for secure computation across organizational boundaries.

Network of connected application icons

Privacy-Preserving Agent Collaboration

Agents reason and compute over sensitive data—including context, memory, credentials, and risk signals—without exposing information, prompts, or business logic.

Financial transaction network diagram

Commerce Protocol Interoperability

A protocol-agnostic binding layer enables negotiation and coordination across agentic commerce standards, including AP2, x402, and the MPP family.

Light beige abstract background

Verifiable Outcome Disclosure

Only the information required by the workflow is disclosed—such as compliance attestations, sanctions results, credential verification, bounded risk scores, or negotiated pricing.

AP3 uses privacy preserving tech.

Animated AP3 graphic

Secure Multi-Party Computation (SMPC)

Agents jointly evaluate functions over private inputs. The inputs never leave the originating agent. The computation happens collaboratively, with no intermediary holding the data.         

Trusted Execution Environment (TEE) attestation

Hardware-rooted proof that the computation ran correctly and was not tampered with. Verifiable without trusting any single party.

AP3 is delivered as an extension to the Agent2Agent (A2A) protocol, compatible with major providers

Google ADK logo
Microsoft AutoGen logo
LangGraph logo
CrewAI logo

FAQs

How do I enable the AP3 SDK in my agent?

What computations does AP3 support?

Is there any additional step required for protocol installation apart from just AP3 SDK install?

When does the agent invoke the AP3 SDK?

In A2A communication, do both agents have to be enabled on AP3 to be able to perform privacy preserving computation?

Where is the AP3 enablement and protocol compatibility declared?

When does the compatibility check happen, to check if both Agents are enabled on AP3 and protocols?

Does the AP3 SDK automatically check for operation compatibility when reading a counterparty's card? Or is that check left to the developer's application logic?

If the agent workflow requires multiple protocols to be used, who defines what protocol is invoked when?

Is there a TypeScript SDK in addition to Python, and are the framework adapters available for both?

In an end to end agent flow, compatibility and handover is required with payment protocols. Is AP3 compatible?

What is the maximum dataset size PSI can handle before latency becomes a problem? Is there a recommended ceiling?

What happens if one party drops out mid-computation, or if there is some timeout? Does the session time out, and is there a retry mechanism?