AgentInteropAudit

AgentInteropAudit is a local-first production auditor for the interoperability boundary between A2A agents and MCP-backed capabilities.

Version: 1.0.0

Why this project exists

A2A v1.0 made agent interoperability substantially more production-oriented: protocol version per interface, supportedInterfaces, A2A-Version, extension negotiation, signatures, multi-tenancy, stronger security declarations, and a formal migration path from v0.3.

A2A's own documentation also defines MCP as a complementary layer: A2A is for agent-to-agent collaboration while MCP is for agent-to-tool/resource connectivity.

AgentInteropAudit audits that boundary.

Capabilities

A2A Agent Card v1.0

• required fields
• supported interfaces
• HTTP+JSON / JSONRPC / GRPC declarations
• protocol-version checks
• HTTPS production endpoints
• media modes
• duplicate interface detection

v0.3 → v1.0 migration

Flags legacy fields such as protocolVersion, url, preferredTransport, additionalInterfaces, and supportsAuthenticatedExtendedCard, plus legacy operation names such as message/send and tasks/get in captured traffic.

Security

• SecurityScheme union consistency
• undefined security requirement references
• skill-level security references
• optional require-auth policy
• extended-card/auth consistency
• Agent Card signature structure
• optional require-signature policy
• possible plaintext credential fields (values are not printed)

Skills and client compatibility

• skill id/name/description/tags
• unique skill IDs
• media compatibility
• client protocol/binding compatibility
• required extension compatibility

Multi-tenancy and traffic

• A2A-Version
• selected interface version
• exact tenant echo
• required A2A-Extensions
• obvious transport credential presence

A2A ↔ MCP bridge governance

Optional bridge manifest:

{"skills":{"web_research":["search_public_web","summarize_sources"]}}

Checks missing skills, missing tools, duplicate tools, object-root MCP schemas, unmapped skills, and orphan tools.

This mapping is explicitly operator-declared. The project does not claim every A2A skill should be an MCP tool; A2A supports richer stateful agent collaboration.

Reports and CI

• console
• JSON
• Markdown
• standalone HTML
• SARIF 2.1.0
• SQLite audit history
• baseline/candidate regression diff
• Mermaid architecture graph

Quick start

python -m pip install .

aiaudit audit   --card examples/agent-card-good.json   --traffic examples/traffic.jsonl   --mcp-tools examples/mcp-tools.json   --bridge examples/bridge.json   --client-profile examples/client-profile.json   --policy examples/policy.json   --min-score 95   --fail-on-error

Generate reports:

aiaudit audit --card examples/agent-card-good.json   --mcp-tools examples/mcp-tools.json   --bridge examples/bridge.json   --json audit.json --markdown audit.md --html audit.html --sarif audit.sarif

Architecture graph:

aiaudit diagram --card examples/agent-card-good.json --mcp-tools examples/mcp-tools.json --bridge examples/bridge.json --output architecture.mmd

Rule families

A2A0xx  Agent Card/interfaces
A2A1xx  security/signatures
A2A2xx  skills
A2A3xx  extensions
A2A4xx  traffic/migration
A2A5xx  A2A↔MCP bridge
A2A6xx  client compatibility

Research basis

See SOURCES.md.

Scope

AgentInteropAudit is not an A2A implementation, not an A2A/MCP conformance suite, not a live scanner, and not a cryptographic JWS verifier. It performs static/passive governance and migration analysis from declarations and captures.

Roadmap

• A2A task/event fixture validation
• v1.0 Part/member-discrimination migration checks
• cursor pagination fixtures
• JWS verification with pluggable key resolvers
• authenticated extended-card diffing
• HAR/OpenTelemetry import
• MCP 2026-07-28 import
• OpenAI Agents SDK / Google ADK / LangGraph adapters
• agent fleet governance dashboard
• trust policy packs
• signed bridge manifests
• JUnit / PR annotations

License

MIT.