MCP
Prompt Injection Through Tool Results
Text your tools return can carry instructions the model may follow. Why you can't filter that away, and how to design a C# MCP server so it can't do much harm.
BLOG
AI engineering in .NET
Short, practical write-ups on building AI features in C#. No roundups, no hype.
MCP
Keep Tool Output Small: Paging and Summaries for the Context Window
Large tool results cost tokens and make agents worse. Cursor paging, trimmed text, aggregate tools and a size budget test for C# MCP servers.
MCP
Read-Only by Default: Designing Safe Write Tools
Write tools let an agent change real data. How to design them in C#: read-only by default, no bulk actions, idempotent writes and user confirmation with MRTR.
MCP
Securing a Remote MCP Server
Lock down a remote C# MCP server with ASP.NET Core: JWT and OAuth metadata, role-based tools, per-user data access, rate limits and an audit trail.
MCP
Tools, Resources and Prompts: Which One When?
MCP gives you three primitives. The real difference is who decides to use them. How to choose between tools, resources and prompts in a C# server.
MCP
Using DI and EF Core DbContext Safely in MCP Tools
How the MCP C# SDK scopes services per request, the lifetime mistakes that break EF Core in tool calls, and the one setting that catches them early.
MCP
Testing MCP Servers: Inspector and In-Memory Tests in xUnit
Test a C# MCP server the way a real client sees it: in-memory client/server tests in xUnit, one real-process smoke test, and the Inspector for manual checks.
MCP
Stateless vs Stateful MCP Servers over HTTP
MCP C# SDK 2.0 made HTTP servers stateless by default. What that changes, how to keep state across tool calls anyway, and when to opt back into sessions.
MCP
stdio or Streamable HTTP? Choosing Your MCP Transport
When a C# MCP server should run as a local stdio process and when it belongs behind Streamable HTTP, plus how to serve the same tools over both.
MCP
Return Errors the Model Can Recover From
How the MCP C# SDK turns exceptions into tool errors, why most of them reach the model as a useless generic message, and how to write errors it can act on.
MCP
Fewer, Bigger Tools: Why 40 Tools Make Your Agent Worse
Mapping every API endpoint to an MCP tool confuses the model. Design C# MCP tools around tasks, not CRUD, and expose only what each client needs.
MCP
Tool Descriptions Are Prompts: Write Them for the Model
The model picks your MCP tools based on their descriptions alone. How to write C# tool and parameter descriptions that lead to the right tool call.
MCP
Your stdio MCP Server Breaks on One Console.WriteLine
A stdio MCP server owns stdout. One stray Console.WriteLine or log line corrupts the protocol. How to route logs to stderr and debug it in .NET.
PROMPTS
Keep Prompts Out of Your C# Code
Store LLM prompts as files, version them automatically with a content hash, and log which prompt produced every answer. A simple pattern for .NET apps.
STRUCTURED OUTPUT
Let the Model Say "I Don't Know": Fighting Hallucinations with Your Schema
LLMs guess when your schema forces them to. Design C# types that allow "unknown", ask for evidence, and verify it in code to catch made-up answers.
MICROSOFT.EXTENSIONS.AI
Switching from Ollama to OpenAI Without Rewriting Your Code
Develop against a local model, run a hosted one in production. How to switch LLM providers in .NET by config, and what silently changes when you do.
MICROSOFT.EXTENSIONS.AI
One Pipeline for Caching, Logging and Tool Calls with IChatClient
Add caching, logging and tool calling to any LLM in .NET with ChatClientBuilder, and learn why the order of your middleware changes what you get back.
STRUCTURED OUTPUT
Stop Parsing LLM JSON by Hand: Structured Output in .NET
Use GetResponseAsync<T> in Microsoft.Extensions.AI to get typed objects from an LLM, handle bad output safely, and keep the model from guessing.
MICROSOFT.EXTENSIONS.AI
IChatClient in .NET: Call a Local LLM from C# and Measure What It Costs
How IChatClient from Microsoft.Extensions.AI works with Ollama and OllamaSharp: first call, tokens, streaming, chat history, ChatOptions, thinking models, DI, logging, caching and switching providers.