2026 · Product, API design
Making a policy engine agent-callable
An MCP server and versioned public contract so AI agents can answer Schedule of Reductions questions with the real engine instead of guessing.
- MCP tools
- 5
- Contract versions
- V1 + V2
- Auth
- Public, read-only
Problem
- People ask assistants about loan limits and get confidently wrong answers, because the assistant has no access to a real, versioned engine.
Constraints
- A public, unauthenticated surface must stay stateless and store nothing about a caller.
- Every answer needs its engine version, policy year, and disclaimer attached so it cannot be laundered into advice.
What I shipped
- An MCP server exposing calculate, list-scenarios, compare, version-compare, and advanced-estimate tools.
- A V2 contract with explicit calculation status, policy decisions, external checks, and warnings on every response.
- A migration and compare surface so consumers can see exactly what changed between versions.
Evidence
- Transparency
- Responses carry engine version, release marker, and an explicit unsupported list.
- Parity tooling
- Public scenario catalogue lets any consumer replay and verify results.
What I would do next
- Publish typed client SDKs generated from the OpenAPI spec.