ADR-004: MCP and A2A Interoperability

Status

Accepted for v5 Phase 1.

Context

DocSentinel must integrate with coding agents, visual workflow builders, and enterprise multi-agent runtimes without coupling the assessment core to each vendor SDK. Agent-triggered work must preserve the same evidence, task, activity, and human-review guarantees as console or REST submissions.

Decision

DocSentinel exposes two complementary protocol surfaces:

• MCP exposes narrow tools and resources. Local stdio remains supported; Streamable HTTP is available at /mcp/.
• A2A 1.0 exposes DocSentinel as a remote specialist agent. Discovery uses /.well-known/agent-card.json; JSON-RPC messages are accepted at /a2a.

REST, MCP, and A2A all call the shared AssessmentService. Protocol adapters must not call the LangGraph orchestrator or mutate task state directly.

Agent submissions always enable collaborative review. An external agent may submit approved documents and read task state, but it may not approve, reject, or bypass a human review gate.

Without AGENT_GATEWAY_TOKEN, protocol endpoints accept loopback clients only. Network deployments require a bearer token, TLS, and preferably an upstream OIDC-aware proxy. MCP DNS-rebinding protection remains enabled, so deployments must explicitly allow trusted Host and Origin values. The public Agent Card and non-secret status endpoint remain discoverable.

Consequences

• Agent platforms integrate through standards instead of vendor-specific core dependencies.
• Security and task behavior stay consistent across entry points.
• The first A2A release accepts references to server-side approved paths. A future document-handle API should replace paths for cross-host workflows.
• OAuth 2.1, signed Agent Cards, task streaming, cancellation, and persistent task storage remain later hardening work.