Co-Authors: ToddSegal, David Soria Parra

Members from various AI Protocols (MCP, A2A) are collaborating on a common AI card standard to benefit the AI Service Ecosystem.

Contact us via GitHub Discussions, Issues, and Pull Requests.

There are multiple evolving standards for communication protocols between AI clients and servers. While these protocols differ in capabilities and technical architecture, there are common ecosystem needs for producers and consumers of services built on top of them. This includes discovery, verifiable metadata, and trusted identity standards.

Without a common discovery standard, we see duplicative and incompatible efforts in the ecosystem, such as registry, marketplace, identity providers, UI and payments extensions, and others. This increases complexity, risk of vulnerabilities, and harms interop within the developer community.

In this repo, we will establish a new open standard for creating, discovering, securing, and serving AI service metadata called an AI Card. This standard will be proposed to be adopted by A2A and MCP following their standard governance processes. Other protocols are encouraged to adopt this standard as well.

Important Note: The AI Card standard does not change the underlying protocol architecture.

The project will define a schema for a secure, tamper-resistant AI Card. This may include items (potential examples):

• Common server metadata, such as publishing company, links to documentation, description, icon, and server version.
• Verifiable metadata, such as accreditations, certifications, reputational scores, and badges.
• Identity metadata, such as Decentralized Identifiers (DID) or Secure Production Identity Framework For Everyone (SPIFFE).
• Custom data defined by individual producers or consumers for their own needs.

Actual properties for inclusion will be debated in PRs.

AI Card will include services information indicating which AI protocols the server supports, their network addresses, and protocol-specific extension metadata such as capabilities.

For example, an AI Card that supports MCP, A2A, and a custom “Foo” protocol, would have entries for each of MCP, A2A, and Foo. Each entry would contain their own specific metadata (such as “MCP tools” or “A2A skills”).

This protocol-specific metadata would NOT be part of the AI Card spec. As noted earlier, the card standard will also allow for arbitrary custom data to be defined by a producer.

As part of the specification, this project would define a standard for how a server can make an individual card, and a catalog of cards, available at a well-known URL for a given domain (e.g. at ./well-known/ai-card.json).

Implementing protocols may support dynamic card creation and discovery through their own methods for scenarios such as providing different card content based on a caller's identity.

Defining this behavior is NOT in scope for the AI Card specification.

The AI Card project should also define a common AI Card Registry standard that provides a universal interface for clients to interact (query, search, CRUDL) with AI card collections.

This is a temporary working repo maintained by the Linux Foundation. Proposals and changes will be made in public in this repo.

This project will be moved to a permanent location a later date with a permanent governance model.

When the specification is finalized, A2A and MCP steering committees will vote on adoption of the AI Card standard, potentially replacing existing protocol-specific standards or proposals.

Protocol-specific cards would be standardized as a common AI Card structure, with protocol-specific properties available in services as noted earlier.

If adopted, MCP and A2A steering committees will recommend duplicative card-adjacent efforts be consolidated, such as Registry, Agent Identity, and others.