OpenSquilla is a microkernel AI agent that runs from a CLI, a Web UI, a desktop app, and chat channels. Every entry point goes through one shared turn loop, so tool dispatch, retries, and decision logging behave the same everywhere. It runs on Windows, macOS, and Linux.
Its central piece is SquillaRouter, an on-device classifier (LightGBM plus ONNX) that scores each turn on length, language, code, keywords, and semantic embeddings, then sends it to the cheapest capable model across four tiers. The decision runs locally, so the prompt does not leave the machine to be classified. The agent also requests extended reasoning and scales the system prompt only for turns the router scores as complex.
A provider layer targets more than 20 LLM backends with primary-plus-fallback selection, including OpenRouter, OpenAI, Anthropic, Ollama, DeepSeek, Gemini, Qwen, Moonshot, Mistral, Zhipu, vLLM, and LM Studio. A high-accuracy mode drafts with several proposer models and fuses the drafts with an aggregator model.
Published benchmark results compare OpenSquilla with OpenClaw and Hermes Agent under one price list. On ClawSWEBench (350 multilingual repair tasks):
| Framework | Model (pool) | Resolve rate | Cost per task |
|---|---|---|---|
| OpenClaw | Opus-4.7 | 77.1% | $3.09 |
| OpenClaw | GLM-5.2 | 74.3% | $0.87 |
| OpenSquilla | GLM-5.2 | 79.4% | $0.95 |
| OpenSquilla | {DeepSeek-V4 Flash, GLM-5.1, GLM-5.2} | 74.0% | $0.44 |
Existing OpenClaw and Hermes Agent state can be imported with a migration command that reports a dry run before applying changes.
Features
- On-device model routing: SquillaRouter classifies each turn locally and routes it across four tiers to the cheapest capable model
- Adaptive reasoning and prompts: extended reasoning and full system instructions apply only to turns scored as complex
- Multi-model ensemble: several proposer models draft in parallel and an aggregator fuses them, with early stopping and per-proposer budgets
- 20+ LLM providers: a provider registry with primary-plus-fallback selection across hosted and local backends
- Persistent local memory: curated Markdown notes searched with SQLite full-text search and sqlite-vec semantic recall, with on-device embeddings or OpenAI and Ollama embeddings
- Layered sandbox: Standard, Strict, and Locked policy tiers, with Bubblewrap on Linux, Seatbelt on macOS, and a native Windows backend, plus a denial ledger that pauses autonomous runs
- Skills and MCP: 15 bundled skills that load on demand, MCP client and server support, and CLI tools to author and publish skills
- Built-in tools: file editing, shell and background processes, git, web search (DuckDuckGo, Bocha, Brave, IQS, Tavily, or Exa), fetch behind an SSRF guard, and office document authoring
- Chat channels: Slack, Telegram, Discord, Feishu, DingTalk, WeCom, Matrix, and QQ share the same turn loop as the Web UI and CLI
- Sessions and scheduling: SQLite-backed sessions and replay, depth-bounded subagents, and a cron scheduler for recurring jobs
- Operator controls: human-in-the-loop approvals for sensitive tool calls and per-turn and per-session token and cost rollups
