🔥 Hot Repo: 9x Cheaper Than OpenClaw, Same Benchmark

One-liner — OpenSquilla is a microkernel AI agent that routes every turn to the cheapest capable model using an on-device ML classifier, hitting near-identical task scores to OpenClaw at 9x lower cost. - Repo: opensquilla/opensquillahttps://github.com/opensquilla/opensquilla - Stars: ⭐ 2,115 +25 today - Language: Python - License: Apache-2.0 --- What It Does OpenSquilla is an open-source AI agent for CLI, Web UI, and 10+ messaging platforms Telegram, Discord, Slack, Feishu, Matrix. Its standout feature is SquillaRouter — an on-device LightGBM + ONNX classifier that scores each turn's complexity using length, language, code patterns, and semantic embeddings, then routes it to one of four model tiers T0–T3. Cheap turns hit budget models; complex turns escalate to frontier ones. The routing decision runs locally — your prompt never leaves the machine to decide which model handles it. Why It's Blowing Up The benchmark data is driving the attention. Against PinchBench 1.2.1 25 real-world tasks, OpenSquilla scored 0.9251 with a mixed-model routing setup versus OpenClaw's 0.9255 on Claude Opus 4.7 alone. Scores are statistically indistinguishable. The cost is not: OpenClaw spent $6.23; OpenSquilla spent $0.69 — 9x cheaper on an identical workload. Token consumption tracked the same pattern — 1.7M input tokens versus OpenClaw's 3M, a 44% reduction — because SquillaRouter also scales the system prompt with task complexity, stripping full instructions down to lightweight variants for trivial turns. The v0.2.0 release May 19 added a further pull: one-command migration from OpenClaw and Hermes Agent configs, including memory, personas, skills, and MCP settings. That targets the large installed base of both competing agents and makes it trivial to run a side-by-side cost comparison before switching. Key Features - SquillaRouter — local ML classifier routes turns across four model tiers; routing decision runs fully on-device - 20+ LLM providers — OpenRouter, OpenAI,