opensquilla/opensquilla

What it solves

OpenSquilla addresses the high cost and complexity of using large language models by providing a token-efficient agent framework. It aims to maximize capability while minimizing budget through intelligent routing and unified interaction across different interfaces.

How it works

The system uses a microkernel architecture where a local model router (SquillaRouter) directs each interaction to the most cost-effective model capable of handling the specific task. It features a single shared turn loop that manages tool dispatch, retries, and decision logging. The agent includes persistent memory, a layered sandbox, built-in web search, and on-device embeddings. It supports a wide range of providers including OpenAI, Anthropic, Ollama, and DeepSeek via a pluggable layer.

Who it’s for

• End users looking for a desktop or CLI-based AI assistant.
• Developers needing a unified agent loop that works identically across Web UI, CLI, and various chat channels (like Slack, Discord, or Telegram).
• Privacy-conscious users who want local routing and the ability to manage data and telemetry.

Highlights

• Intelligent Routing: Uses an on-device router to select the cheapest capable model.
• Multi-Interface Support: Consistent behavior across Web UI, CLI, and chat channels.
• Extensive Provider Support: Connects to over 20 LLM providers including OpenRouter and local models via Ollama.
• Rich Feature Set: Includes persistent memory, web search, and sandboxed execution.
• Flexible Deployment: Available as a desktop app, a CLI tool, or via source installation.