OpenSquilla Launches Open Source AI Runtime With ML Routing And Secure Sandboxing

OpenSquilla has released the first public version of its self-hostable, open-source AI agent runtime, positioning it as a cost-efficient alternative to conventional AI agent stacks for long-horizon enterprise workloads. Released under the Apache-2.0 licence, the Python 3.12+ framework is available for self-hosting on GitHub.

The project's primary focus is reducing unnecessary AI token expenditure. OpenSquilla claims its coordinated routing and optimisation stack can lower token spending by 60–80% compared to flat single-model deployments. Built-in quota hooks and per-call cost tracking are designed to automatically detect and throttle overspending.

In a local benchmark, three prompts processed a combined 279,762 tokens at a total session cost of $0.0094. Around 222,848 tokens — nearly 80% of all input tokens — were served from cache through context reuse across sessions. That's the kind of number that makes enterprise procurement teams sit up (or at least stop groaning).

The runtime uses an ML classifier that evaluates request complexity using message length, code blocks, keyword patterns, and embedding-based semantic features. Simpler tasks are routed to lower-cost models, while deep reasoning is disabled for lightweight prompts to reduce compute overhead.

According to the official GitHub repository, the framework combines smart routing, persistent memory, a secure sandbox, built-in web search, and local embeddings under a single model loop. Every entry point — Web UI, CLI, and chat channels — runs through a shared TurnRunner, and a pluggable provider layer lets it speak to OpenRouter, OpenAI, Anthropic, Ollama, DeepSeek, Gemini, Qwen/DashScope, and roughly twenty other LLM providers without changes to your code or config schema.

OpenSquilla also introduces a four-tier cognitive memory architecture comprising working, episodic, semantic, and raw memory layers, alongside vector-semantic and BM25 retrieval. Local ONNX inference keeps embeddings on-device, which means the physical latency of network calls for embedding generation disappears entirely.

On security, the framework uses syscall-level isolation through Bubblewrap on Linux and Seatbelt on macOS, alongside policy-based execution controls and prompt injection protections. Its microkernel-style architecture further enables lightweight plugin creation without mandatory SDKs or manifest files. This is a notable departure from the Docker-heavy approach that has become standard in the AI agent space.

Independent reporting from Open Source For You corroborates the timeline and scope of the changes, confirming the Apache-2.0 licensing and the 60-80% token savings claim.

Installation paths vary by user type. New users can download the preview release package from the GitHub Releases page and extract it to a writable folder. Command-line users install from source, while developers can develop from source directly. The portable zip does not install a global opensquilla command — for a terminal where opensquilla commands work, users must run OpenSquilla Shell.cmd or use the extracted folder through .\opensquilla.cmd.

The physical experience of running OpenSquilla differs from typical AI agent deployments. The launcher opens onboarding before the gateway starts. On first run, users choose a provider and paste the requested keys; later starts let them review or change the config. Then they open http://127.0.0.1:18790/control/ in their browser. The terminal window must stay open — closing it stops the gateway entirely.

Whether the 60-80% savings hold up in production environments with real enterprise workloads remains the real question. Benchmarks are one thing; actual deployment with unpredictable user queries is another. The framework's architecture suggests it could work, but the token savings depend heavily on how well the ML classifier routes requests in the wild.

For now, the preview packages serve as the recommended public distribution channel for validating installation, onboarding, the local gateway, and the Web UI before the stable 0.1.0 release. Developers wanting to skip the bundled router can set OPENSQUILLA_INSTALL_PROFILE=core, though that defeats the primary value proposition.

Whether users actually pay for the convenience of this routing layer, or simply build their own, remains to be seen. The open-source nature means the barrier to entry is low, but the barrier to meaningful cost savings might be higher than the benchmarks suggest.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn