What the MPAI Store is and does
MPAI — "Moving Picture, Audio and Data Coding by Artificial Intelligence" — is a standards organization that builds AI applications from standardized, swappable building blocks instead of monolithic systems. The MPAI Store is the ecosystem's trusted distribution hub.
The building blocks
| Term | What it is |
|---|---|
| AIM — AI Module | The atomic processing component, with a defined function and a standardized interface (specified input/output). Any AIM can be swapped for another implementation with the same function and interface — this is MPAI's notion of interoperability. A composite AIM contains SubAIMs. |
| AIW — AI Workflow | An assembly of AIMs wired together to perform a complete use case (e.g. answering a spoken question about an image). |
| AIF — AI Framework (MPAI-AIF) | The standardized runtime that executes AIWs and AIMs. It provides a Controller plus services for inter-AIM communication, shared storage, and access to the MPAI Store. |
| Metadata | Each AIW/AIM ships a JSON metadata document describing its identity, ports, sub-AIMs, topology, concrete implementations, and resource policies. That document is what this Store validates, rewrites, and publishes. |
The Store's role
Under the MPAI-GME (Governance of the MPAI Ecosystem) standard, the Store is three things at once. This is the role the standard defines — not a description of what this registry performs; see the box below for that.
- Registration Authority — registers developers ("Implementers") and issues each a unique Implementer ID. Performed here.
- Verification gate — checks submissions before distribution, and tests their claimed conformance to the relevant MPAI standard. Partly performed here: this registry validates metadata against the schema and authenticates the implementer. It does not test conformance.
- Distribution catalogue — publishes approved implementations, labeled with their Interoperability Level, for consumers to find and run. Performed here as a catalogue only — implementers host their own artifacts; this registry never serves them.
- We do not perform a security review of the implementation code, and we do not host, build, run, or scan the artifact. We store its URL and its SHA-256 and sign a statement binding the two.
- We do not run conformance testing or performance assessment. No independent Performance Assessor is involved.
The five ecosystem roles
These are the roles as MPAI-GME defines them, for the ecosystem as a whole. The "MPAI Store" row describes what a Store is specified to do — see the box above for what this registry does.
| Role | Responsibility |
|---|---|
| MPAI | Develops the standards (Technical, Reference Software, Conformance, Performance). |
| Implementers | Build AIM/AIW/AIF implementations and submit them to the Store. |
| Performance Assessors | MPAI-appointed bodies that assess an implementation's performance (reliability, robustness, fairness, replicability). |
| MPAI Store | Registers Implementers, verifies security, tests conformance, labels by Interoperability Level, and distributes. This registry performs registration, schema validation and labelling only — it does not verify security, test conformance, or host artifacts. |
| Users | Download and run implementations; may report Experience Scores. |
The lifecycle
- Onboard. A developer registers and is issued an Implementer ID (/register).
- Submit. The Implementer sends the AIM/AIW metadata (and, in the full flow, the package, a fingerprint, and the URL where binaries are hosted) (/submit).
- Assess. The Store checks the submission's structure against the MPAI metadata schema and the implementer's credentials before it can be published.
- Assign & publish. The Store assigns an Implementation ID, stamps the metadata, rewrites the
ImplementationsandResourcePoliciessections, signs a trust statement, and publishes — returning the rewritten metadata to the developer. - Consume. A consumer's SCI (Store Consumer Interface) searches the Store, reads the Implementation ID, fetches the metadata and signed statement, verifies them, downloads the AIMs, and a Controller instantiates and runs them inside an AIF.
Interoperability Levels (the trust signal)
MPAI labels each published implementation with one of three levels; trust increases from 1 to 3. The "what is verified" column describes what each level means in the MPAI standard — not what this registry performs. As above, this registry issues Level 1 only, on the basis of schema validation and implementer authentication.
| Level | Meaning | What is verified |
|---|---|---|
| Level 1 | Implementer-specific, but runs in a conformant MPAI-AIF. | Security verified; conformance to MPAI-AIF tested. |
| Level 2 | Conforms to a published MPAI Application Standard (e.g. MPAI-MMC). | Security verified; conformance to that standard tested. |
| Level 3 | Level 2, plus validated by a Performance Assessor. | Additionally assessed Reliable, Robust, Fair, Replicable. |
Worked example: MMC-AMQ
MMC-AMQ ("Answer to Multimodal Question") is an AI Workflow in the MPAI-MMC (Multimodal Conversation) standard: it takes a question as text/speech together with an image and produces an answer. Its top-level AIMs are:
The composite AIM with three SubAIMs is MMC-TIQ — a good worked example of the shape your metadata takes, because it exercises SubAIMs, ports and a topology all at once. Browse the registry to see what has been published so far.