MyCodingAgent is an Ollama-first, .NET-focused development system that turns LLMs into reliable engineering agents instead of unpredictable code generators.
It acts as a structured layer between your codebase and local LLMs, enforcing correctness, consistency, and architecture through Roslyn-powered validation and controlled generation pipelines.
Runs entirely on local models via Ollama. No cloud dependency. No hidden costs. Full control over models, context, and behavior.
MyCodingAgent is built exclusively for the .NET ecosystem:
- Blazor (frontend)
- Web API
- Console apps & workers
- Service Bus / microservices
- Docker-based environments
- Infrastructure as Code
This focus allows deep integration instead of shallow generalization.
All generated code is validated through the Roslyn compiler pipeline:
- Prevents invalid code from entering the system
- Enables semantic analysis instead of string-based guessing
- Forces LLM output into compilable, structured reality
MyCodingAgent is not “AI that writes code”.
It is:
- A controlled code generation system
- A context-aware prompt engine
- A deterministic orchestration layer for LLM workflows
Instead of letting an LLM “figure things out”, MyCodingAgent:
- Defines the task
- Limits the context
- Generates targeted output
- Validates via Roslyn
- Iterates until correct
Rebuild system state from a minimal starting point using:
- Event sourcing
- Fully reproducible generation steps
- No hidden mutations
- Built-in task/subtask system
- Local version control integration
- Prompt generation based on actual project state
- Retrieval-Augmented Generation tailored for .NET
- Roslyn analyzers feeding structured insights into LLM context
- Smarter, safer code generation
- Distributed agent system
- Modular capability providers
- But first: a strong standalone core
LLMs are powerful, but unreliable by default.
MyCodingAgent exists to:
- Remove randomness
- Enforce structure
- Make LLMs usable for real engineering work
But also because there are no tools like this that use Ollama that I know off. And I could use the experience.
Work in progress. Built for experimentation, iteration, and pushing the limits of local AI-driven development. I've booked very good results with testing but it's not ready jet. Most likely there are errors and deadlocks as I'm still building the foundation right.