0003 — Bedrock is an optional backend, not a hard dependency

• Status: accepted
• Date: 2026-05-15

Context and problem statement

Several tools (suggest_services, generate_architecture_diagram, review_well_architected) could benefit from LLM reasoning to produce richer, less keyword-driven output. We have to decide whether that is required or optional.

Considerations:

• The server is meant to be embedded in an MCP client that is already an LLM. Adding a second LLM call inside the server can be redundant.
• Requiring AWS credentials and a Bedrock model would raise the barrier to first-run dramatically.
• The deterministic, keyword-based implementation is auditable — anyone can read the catalog and the rules in a few minutes.

Decision

Tools are deterministic by default. Bedrock is an optional backend that can be enabled per tool via the bedrock extras and an env-var-toggled flag. When disabled (the default), every tool ships a useful, transparent response without making any AWS API call.

Alternatives considered

• Bedrock as the primary backend. Rejected — raises barrier, slower iteration, harder to test.
• No LLM ever. Rejected as a long-term stance — some tasks (free-form architecture review) are genuinely LLM-shaped.

Consequences

• The first-run experience is pip install + run. No AWS setup needed.
• Bedrock integration is a clean, optional layer we can ship later without breaking the deterministic path.
• The keyword-based catalog and rules must stay readable and well-tested — they remain the reference implementation.