OpenSquilla Deep Dive: How Local Routing Cuts Token Costs to 1/9
In one sentence
OpenSquilla reduces agent LLM call costs from a higher baseline to $0.68 with almost no quality loss. The core trick: a local small model acts as a judge — simple tasks go to cheap models, complex tasks get top-tier models.
---
Key figures 📊
| Dimension | Value | |------|------| | License | Apache 2.0 | | Current version | 0.3.1 | | GitHub Stars | ~2,000+ | | Supported models | 20+ providers | | Cost reduction | ~90% | | Routing decision latency | Zero token consumption (runs locally) |
---
SquillaRouter: the local judge model
Uses a LightGBM + ONNX classifier. Inputs: length, language, code presence, keywords, and semantic embeddings. Output: a T0–T3 difficulty tier.
Key point: the judge runs entirely locally — prompts never leave your machine to ask a big model how hard the task is.
Benchmark: across 25 tasks, pure Claude Opus scored 0.9255; OpenSquilla's mixed routing scored 0.9251 at $0.68, cutting cost by ~89% for a 0.0004 score difference.
---
Four-layer memory: long-term experience for agents
- Working memory: current conversation context
- Short-term memory: recent summaries
- Long-term memory: Markdown notes
- Archival memory: SQLite full-text search + semantic recall
- Complex skills get skipped by the AI
- Errors vary each run; sometimes results are fabricated
- Rules are clear, but the agent doesn't follow the sequence
- Mixed workloads of frequent simple + rare complex tasks (support agents: 80% FAQ + 20% technical issues)
- Budget-sensitive production environments (need Opus-level capability but can't afford full-price calls)
- Multi-platform deployment (one agent serving Web, Slack, Discord, Feishu, etc.)
- Local-first security requirements (routing decisions and embeddings stay fully local; no data leaves the machine)
Context is automatically compressed when full, and persists across sessions. Incremental transfer avoids resending history, reducing repeated tokens by ~90%.
---
MetaSkill: fixing disobedient agents
Common agent frustrations:
MetaSkill applies meta-level control over skills, ensuring execution order and completeness — solving both "won't run" and "won't behave".
---
Three-tier security sandbox
| Tier | Linux | macOS | Windows | |------|-------|-------|---------| | Standard | ✅ | ✅ | ✅ | | Strict | ✅ | rendering only | ❌ | | Locked | ✅ | not yet | ❌ |
Linux has full Bubblewrap isolation; macOS/Windows support is still catching up.
---
Limitations ⚠️
1. The routing model depends on training data — brand-new task types may be misjudged in difficulty 2. The cost reduction is scenario-specific; gains shrink for fully complex workloads 3. Ecosystem maturity lags behind OpenClaw — stars are growing fast but real-world cases are few 4. Windows/macOS sandboxing is incomplete
---
Who should use it
Key takeaways
1. Core value: local intelligent routing cuts costs to 1/9 at equivalent quality 2. Technical highlights: zero-token LightGBM routing, four-layer memory, MetaSkill execution control 3. Best fit: scenarios where simple tasks make up 60%+ of traffic 4. Risks: cost data is scenario-specific; young ecosystem; cross-platform sandbox incomplete
---
Sources: official GitHub repository and docs; third-party reviews (CSDN/xmsumi.com). Research date: 2026-06-05