Microkernel agent swaps LLMs mid-chat to cut costs

What it does

OpenSquilla is a microkernel AI agent built around a single shared turn loop that powers its Web UI, CLI, and chat channels. A local component called SquillaRouter selects the cheapest available model for each individual turn from a pluggable roster of providers—OpenRouter, OpenAI, Anthropic, Ollama, DeepSeek, Gemini, Qwen, and 20-plus others—while persistent memory, on-device embeddings, a layered sandbox, and built-in web search all operate inside the same loop.

The interesting bit

The routing decision runs locally via ONNX and LightGBM, meaning your machine decides which remote API to call before spending the token. That keeps the budget-optimization logic on-device even when the heavy lifting is outsourced.

Key highlights

• SquillaRouter routes each conversational turn to the cheapest capable model
• Identical tool dispatch, retries, and decision logging across Web UI, CLI, and chat channels (Slack, Discord, Telegram, Matrix, Feishu, and others)
• Memory and embeddings run on-device rather than in the cloud
• Built-in layered sandbox and web search are core loop citizens, not bolt-ons
• Provider swaps require no changes to code or config schema

Caveats

• Windows portable builds are unsigned and require administrator launch with a SmartScreen bypass
• The quick terminal install path on Windows omits the Visual C++ runtime required by SquillaRouter’s ONNX dependency, falling back to direct single-model routing until installed
• Advanced channel features like Matrix end-to-end encryption and PDF generation require separate opt-in extras

Verdict

A solid experiment for developers running multi-provider agent workloads who want per-turn cost optimization. Less compelling if you prefer a fully managed, single-model SaaS agent with zero local footprint.