Cyréna is an experimental offline-first engineering assistant that runs alongside your IDE, not inside it.
It operates as a structured agent inside real projects and helps with:
- building features
- repairing compile errors
- enforcing architecture
- maintaining technical documentation
- preserving long-term project knowledge
This project started as a personal experiment to build the assistant I wanted for my own workflow.
It is evolving into something larger.
It contains important safety notes.
Here is Cyréna running a real prompt → build → repair loop:
This is not code generation —
this is an engineering agent working inside a project.
Cyréna is not a chatbot.
It is a constrained engineering agent that:
- reads a project plan
- inspects real files
- makes minimal edits
- runs iterative repair loops
- persists structured knowledge
- behaves like a disciplined developer
The goal is predictable AI behavior inside real codebases.
Cyréna is designed as a platform, not a single-language tool.
Currently supported:
- 📦 .NET
- C# Class Library
- Blazor App
- C# MVC App
- C# MVC Library note: It follows the structure of MVC App
- 🔌 Arduino IDE firmware projects
- 🔌 PlatformIO firmware projects
- Arduino & ESPIDF
Each domain has its own constraints and prompts, but shares the same core architecture.
This proves the system is extensible.
API References are structured technical documents generated by the AI and stored with the project.
They describe:
- APIs
- architecture rules
- contracts
- module behavior
- integration patterns
They act as persistent technical memory for the agent.
The AI reasons from real documentation grounded in real code.
This allows Cyréna to scale beyond context window limits.
Cyréna enforces strict project structure rules.
The agent cannot invent random folder layouts or escape the project boundary.
This prevents entropy and keeps generated code maintainable.
For Arduino projects, the flat sketch model is enforced.
For .NET projects & PlatformIO, structured folders are enforced.
Different domains, same discipline.
Every action runs through a controlled loop:
- Read project plan
- Search API References
- Review project notes
- Inspect relevant files
- Implement minimal change
- Build / validate
- Repair if needed
- Summarize
- Persist knowledge
Chat memory resets each iteration.
Structured knowledge survives.
This avoids long-session drift.
- .NET 10
- Windows / Linux / macOS
- Ollama or OpenAI
- Recommended: ≥16k context models
Hardware requirements scale with model size.
Documentation is evolving — thank you for your patience.
- Run Cyréna
- Open settings (gear icon)
- Add an Ollama connection
- Set model + tokens + context
- (Optional) Add Tavily API key
- Open settings
- Enter API key + model ID
- Save
- Return to main screen
- Click +
- Select project type
- Fill in details
- Submit
Baseline dev system:
- RTX 3060 (12GB)
- 48GB RAM
- Ryzen 7 8700F
Test model:
gpt-oss:20b- 8k tokens
- 16k context
Bigger models need more VRAM + RAM.
Cyréna stands on the shoulders of great projects:
Cyréna is under active development.
It is an autonomous agent capable of modifying real files.
Use version control.
Use backups.
Review all changes.
This is not production-hardened software yet.
- Incorrect overwrites
- Imperfect repair loops
- No full rollback yet
- Intent may drift
Always review AI edits.
- Diff keeper
- Revert mechanism
- Safe execution guardrails
- Hard command sandbox: note only supports dotnet build for now
Until complete:
👉 Treat Cyréna as experimental
👉 Expect mistakes
👉 Verify output
Cyréna is being built toward:
- disciplined AI behavior
- reversible changes
- predictable edits
- engineering safety
- long-term project memory
You are seeing the system mid-evolution.
Cyréna is experimental software.
It may introduce errors or data loss.
You are responsible for backups and review.
Provided as-is. Use at your own risk.