OpenSquilla: Token-Efficient Open Source AI Agent with Smart Model Routing

Why Another AI Agent?

I've tested a lot of AI agents. Some are smart but expensive. Some are cheap but useless. A few are both — just don't work with anything beyond OpenAI.

OpenSquilla does something I haven't seen done well before: it routes each request to the cheapest model that can actually handle it. Think about that — instead of blasting every query through GPT-4, it checks if Gemini Flash or a local Ollama model can do the job for pennies.

What Makes OpenSquilla Different?

It's built around a microkernel architecture — a fancy way of saying everything runs through the same core loop whether you're using the Web UI, the terminal, or a chat channel. That means your tools, memory, and routing behave the same everywhere. No surprises.

1. The SquillaRouter

This is the brain. It's an on-device model router that learns which model is best for which task over time. Start with a simple config, and it gets smarter as you use it. The router supports 20+ LLM providers — OpenAI, Anthropic, Ollama, DeepSeek, Gemini, Qwen, and more — all without changing a single line of config.

2. Persistent Memory That Actually Works

Most agent "memory" is a glorified chat history file. OpenSquilla gives you persistent memory with sandboxed execution — your agent remembers context across sessions, and the sandbox keeps experiments contained. On-device embeddings mean your vector search doesn't hit an external API.

3. MCP Protocol Support

OpenSquilla speaks the Model Context Protocol (MCP) natively. That means it can plug into the growing MCP ecosystem of tools and servers without custom adapters. Need a database tool? An API connector? Read the MCP server docs and you're done.

Where It Shines

I've been using it for research tasks where I need multiple models to fact-check each other. The routing means I spend ~60% less on API costs compared to routing everything through a single frontier model. For coding tasks, it connects to code repositories and handles multi-step workflows through the shared loop.

Conclusion

OpenSquilla isn't trying to be another agent framework. It's a production-ready agent that treats token cost as a first-class concern. If you're tired of burning API credits on simple queries or want a single agent that works across CLI, Web, and chat, this is worth your time.

Who it's for: Developers building multi-provider agent workflows. Teams that want consistent behavior across interfaces. Anyone tired of paying GPT-4 prices for "what's the weather" queries.

Who it's not for: People who want a no-code agent builder. If you're not comfortable with a config file and a terminal, start with something simpler.