An MCP server for a wealth-advisory rules engine
Taffrail runs the rules engine behind Advice-as-a-Service for wealth and financial advisory teams. Wrapped the rules engine and the OpenAPI surface in two MCP servers so AI agents and developer tools (Cursor, Claude Desktop, Windsurf) can query live financial advice — not hallucinated approximations of it.
the problem
Taffrail’s value sits in a rules engine — a hardened, compliance-aware system that turns user context into specific financial advice. The web product made that engine available to advisors. Agents could not get to it. An LLM asked a wealth-planning question would hallucinate an answer instead of calling Taffrail’s actual logic, and developers integrating with Taffrail’s API had to round-trip through docs and Postman every time they wrote a query.
Two audiences. One missing surface.
what shipped
Two MCP servers, both pointing at the same underlying truth.
- Rules MCP —
mcp.taffrail.com. Lets AI agents query Taffrail rules, run variable lookups, and retrieve advice on behalf of a user. Agents now route financial questions through the engine instead of guessing. - Docs MCP —
docs.taffrail.io/mcp. Exposes the OpenAPI spec and developer guides to Cursor, Claude Desktop, and Windsurf. Developers building integrations get live, version-correct API answers from inside their editor — no doc-tab juggling, no training-data drift.
Configuration is a single JSON block per editor. Verification is a single question: ask the assistant something Taffrail-specific and watch it pull the live spec.
what changed
The engine became reachable from where agents and developers actually live. The “advice as a service” framing got a second meaning — advice as a service for AI applications, not just human advisors. And the documentation-MCP pattern turned out to be a quiet force multiplier: every dev tool that adds MCP support now ships Taffrail integration for free.
the lesson
If you have a real engine — rules, data, judgment — make agents call it. Don’t let them guess. MCP is the cheapest way to put your real logic where the models are already shopping for context.