Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-09-22. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Free (Apache-2.0) | 2026-09-22 |
One-liner: OpenSquilla is a token-efficient AI agent with a microkernel design: a local model router sends each turn to the cheapest model capable of handling it, inside one shared turn loop.
Verified facts (checked 2026-09-22)
| Attribute | Value | Source |
|---|---|---|
| GitHub stars | 7,037 (570 forks) | TokenRhythm/opensquilla via GitHub API, 2026-09-22 |
| License | Apache-2.0 | GitHub API, 2026-09-22 |
| Stable release | 0.5.4 | repository README |
| Repository created | 2026-05-06 | GitHub API |
| Last push | 2026-09-21 | GitHub API |
| Runtime requirement | Python 3.12+ | repository README |
| Platforms | Windows, macOS, Linux (desktop installers for macOS/Windows) | repository README |
What it actually does
The core idea is routing discipline. A local component called SquillaRouter decides, turn by turn, which model should handle the request, with cost as an explicit routing dimension. Persistent memory, a layered sandbox, built-in web search and on-device embeddings all run through a single shared turn loop, so the Web UI, CLI and chat channels behave identically: same tool dispatch, same retries, same decision logging.
The provider layer is provider-agnostic — TokenRhythm, OpenRouter, OpenAI, Anthropic, Ollama, DeepSeek, Gemini, Qwen/DashScope and 20+ others plug in without changing your config schema. Installation is flexible: desktop DMG/EXE installers (OpenSquilla-0.5.4-mac-arm64.dmg, OpenSquilla-0.5.4-win-x64.exe), a quick terminal install via uv, or a Git checkout for contributors. The team has also published technical reports on the routing approach (an arXiv paper on agentic routing and an aiXiv/ChinaXiv report).
What the marketing doesn’t say
The router is not free of operational friction. On Windows, its bundled ONNX runtime needs the Visual C++ runtime; the quick terminal install path does not install it automatically, and a DLL load failed log means a manual fix. On macOS terminal installs, the LightGBM runtime can need brew install libomp. In both cases OpenSquilla falls back to direct single-model routing until the dependency is present, so the agent still runs. If you don’t want the router at all, a core install profile or a --router disabled flag turns it off.
Widely shared claims about fixed token-saving percentages are not adopted here without primary measurement: routing can reduce spend, but your savings depend entirely on your model mix and traffic.
Who should use it
Cost-sensitive developers and small teams running multi-provider agents who want routing logic on their own machine under Apache-2.0, and who are comfortable with a Python 3.12+ environment. If you prefer a fully managed SaaS where routing is someone else’s operational concern, this isn’t that product.
Repository stats, license text and install requirements verified against primary sources on 2026-09-22. This entry follows our verification policy: every claim here was checked against primary sources, not copied from other reviews.