Building Pebbles: A Journey in AI Tooling

I built Pebbles, a .NET CLI AI assistant, and learned a lot along the way. Here's what happened. Making It Work The first big challenge was multi-provider support. I started with Alibaba Cloud's DashScope, but wanted to try other models. Each provider has its own API format, authentication, and quirks. OpenAI uses Bearer tokens, Anthropic uses different headers, and they all have different features. I built a unified interface to hide the mess and created an interactive setup flow so new users don't have to edit config files. I also added 173 tests because supporting multiple providers means testing all of them. Then I gave Pebbles actual capabilities. It could chat, but it couldn't do anything. I added three tools: run shell commands, write files, and browse directories. This required serious safety work: blocking dangerous commands like rm -rf / , validating file paths, and creating automatic backups before every change. I used Polly for retry logic because networks fail and files ge