AI
MCP Deep Dive, Part 5: Designing Custom MCP Tools Your Agents Actually Use Right
Most MCP tools mirror your REST API and confuse the model. Here's how to design task-shaped tools an agent picks right, calls cleanly, and trusts.
15 min read11
Long-form tutorials and engineering notes on .NET, React, AI/ML, and SaaS architecture.
Most MCP tools mirror your REST API and confuse the model. Here's how to design task-shaped tools an agent picks right, calls cleanly, and trusts.
The server exposes tools; the client drives them. Here's a production MCP client — discovery, the agent loop, multi-server routing, and reconnects.
A demo MCP server is a weekend; a production one needs typed tools, real error handling, pagination, and health checks. The Mattrx build, end to end.
MCP has a host, clients, servers, three primitives, and a capability handshake. Here is the full architecture, mapped onto Mattrx's real Azure setup.
Every AI agent you build needs every backend you own — that is N×M glue code. MCP turns it into N+M. Here is why, with real Mattrx production numbers.
Naive AI review drowns devs in false positives until they ignore it. Here's the context-aware, adversarially-verified pipeline we run on every Mattrx PR.
Chunk-and-embed RAG hits a wall at scale. Open Knowledge Format (OKF) feeds a context engine structured, governed knowledge — the Mattrx rebuild with code.
Most enterprises bolt AI onto a backend built for CRUD. We rebuilt Mattrx around nine AI-native layers in production. Here is the blueprint, with code.
Part 6: AI & data governance as a control plane — classification at ingestion, entitlement-aware retrieval, consent & purpose limits, data lineage, and policy-as-code in C# + Python.
Part 5: multi-tenant context engineering — tenant-scoped retrieval, per-tenant prompts and caches, noisy-neighbor fairness, model routing, and per-tenant cost attribution in C# + Python.
Part 4 : the enterprise design that makes GenAI production-grade — eval gates, injection + PII defense, cost control, and tracing.
Part 3: multi-agent architecture that survives production — a supervisor + specialized agents (C# orchestration, Python agent graph) with bounded loops.