Architecture Intelligence Platform

What Architecture Intelligence Platform does

Architecture Intelligence Platform MCP server helps coding agents inspect service relationships without reconstructing architecture from source code alone. It reconciles declared API and event relationships with dependencies observed during a selected runtime window.

The result is a snapshot-bound answer with qualifications such as CONFIRMED, OBSERVED_ONLY, and NOT_OBSERVED_IN_WINDOW. A declared relationship seen in telemetry is confirmed; a runtime dependency with no corresponding declaration is observed only; a declaration missing from the selected window is reported as not observed, rather than automatically labeled dead or unused.

Each claim includes evidence references, and unresolved identity is represented as a limitation instead of being replaced with a guess. The tools are read-only.

How it works

Architecture Intelligence Platform MCP server combines declared architecture inputs, including AsyncAPI examples or overlays, with OpenTelemetry observations. The supplied demonstrations run the platform with Docker Compose, Neo4j, and an OpenTelemetry Collector. A request is evaluated against a particular snapshot and runtime window, so related queries can be tied back to the same evidence state.

The MCP interaction can be driven through HTTP and JSON-RPC. The five-minute demo starts the supporting components, imports an example architecture, seeds timestamp-frozen telemetry, and calls tools/list, get_architecture_drift, and get_evidence. Repeated runs with the frozen data produce the same qualifications and snapshot identifier.

Setup and configuration

The repository provides a five-minute local path that requires Docker with Compose, curl, and jq. It does not require an LLM API key for that demo. The documented sequence clones the repository, copies .env.example to .env, and runs examples/runtime-demo/mcp-demo.sh.

Additional replayable demonstrations cover Quarkus Super Heroes and Pitstop. Their local scripts require Docker and curl; the documented runs do not require building Quarkus or .NET, running Kafka or RabbitMQ, or supplying a model key. The Pitstop hosted read-only endpoint is described as requiring no credentials.

Tools and capabilities

Architecture Intelligence Platform MCP server exposes read-only tools for:

• Finding discrepancies with get_architecture_drift, including undocumented observed dependencies and declared relationships absent from the selected window.
• Listing a service's complete dependency set with get_service_dependencies, including relationships that are confirmed rather than drift findings.
• Resolving claim provenance with get_evidence from evidence references at the matching snapshot.
• Returning limitations when identity or another architectural fact cannot be established safely.

The demonstrations show operation-level calls, deployment mappings, published topics, named queues, and evidence from AsyncAPI and OpenTelemetry. Exact coverage depends on the supplied declarations, observations, mappings, and runtime window.

Limitations and notes

Architecture Intelligence Platform MCP server does not claim facts it cannot support. A consumer may remain unresolved when evidence does not identify one service, and the answer can be partial. The Pitstop example explicitly retains the need for source inspection when the system has no payload or field-level knowledge.

The platform reports discrepancies rather than matches in drift results. A relationship not observed in one window is not proof that it is unused. The README also notes that four coding-agent clients were qualified against version v0.4.2, but that qualification has not been repeated since. The examples include a gRPC call that the platform cannot yet answer and label it outside the answer.