Opensquilla
OpenSquilla — Token-Efficient AI Agent with same budget, higher intelligence density
Last verified:
What is Opensquilla?
OpenSquilla is a token-efficient microkernel AI agent runtime designed to reduce enterprise token costs by 60-80% through intelligent model routing, multi-tier memory, and secure syscall-level isolation. It is a self-hostable, open-source Python 3.12+ framework released under Apache License 2.0, available on GitHub. The tool targets developers and teams running AI agents for sustained, long-horizon enterprise workloads where token bills compound across sessions.
Key features include smart ML-based model routing that evaluates request complexity using message length, code blocks, keyword patterns, and embedding-based semantic features to route simple tasks to cheaper models. It has a four-tier cognitive memory architecture (working, episodic, semantic, raw) with vector-semantic and BM25 hybrid search, local ONNX embeddings, hot memory promotion, temporal decay, and automated
Opensquilla pricing
Pricing model: Freemium
OpenSquilla is fully open source under Apache License 2.0 and free to self-host from GitHub. There is no paid subscription for the software itself. Running LLMs still requires token credits from providers (OpenAI, Ollama, etc.) which users must supply separately. The website offers a limited-time free token credit trial with zero credit card required to test the cost savings. Users can claim free tokens via a 10-second signup or use the open-source version on GitHub with their own tokens. The portable version includes bundled CPython runtime at no extra cost.
Opensquilla pros
- 60-80% token cost savings through ML routing and multi-strategy optimization
- Microkernel architecture with ultra-compact ~100 line core orchestrator
- Four-tier cognitive memory mirroring human cognition (working, episodic, semantic, raw)
- Syscall-level sandbox isolation without Docker dependency
- Built-in web search with Brave and DuckDuckGo integrations
- Local ONNX embeddings keep data on-device, optional OpenAI/Ollama swap
- 10+ communication channels built-in (Slack, Discord, Telegram, Teams, Matrix, etc.)
- Hot memory promotion auto-bubbles frequently recalled memories to top
- Memory Dream consolidation automatically structures knowledge every 24 hours
- Three-tier security policy system (Standard, Strict, Locked) with human review gates
- 5-line duck-typed plugin system with no base class or SDK required
- Temporal decay fades dated memories while evergreen items stay sharp
- Auto-rollback on errors and fully async pipeline orchestrator
- One-click migration from OpenClaw or Hermes Agent frameworks
- Portable Windows ZIP with bundled CPython, no separate Python install needed
- Decision logs store hashes not raw text for compliance-friendly observability
- Reasoning depth tiers disable deep thought billing for simple queries
- On-demand skills only load capabilities needed for current task
- Denial ledger auto-pauses AI after three consecutive security rejections
- Quota hooks enable auto-throttling when overspend detected
Opensquilla cons
- Requires Python 3.12+ which may not be available on older systems
- Preview builds are unsigned requiring admin rights and SmartScreen bypass on Windows
- Must keep terminal window open or gateway stops running
- ML routing dependencies (ONNX Runtime, LightGBM) add installation complexity
- No native mobile app, only web and communication channel integrations
- Self-hosting requires manual setup and configuration of provider API keys
- Bubblewrap sandbox only available on Linux, Seatbelt on macOS (no Windows sandbox)
- Memory Dream consolidation runs on fixed 24-hour schedule, not configurable
- Free token credit is limited quantity first-come-first-served
- Web UI stability issues reported in long-running sessions before v0.2.1
Frequently asked questions about Opensquilla
What is OpenSquilla?
OpenSquilla is a token-efficient microkernel AI agent runtime designed to cut enterprise token costs by 60-80% through intelligent ML-based model routing, four-tier cognitive memory, and secure syscall-level sandbox isolation. It is self-hostable, open source under Apache 2.0, built on Python 3.12+, and available on GitHub.
How does OpenSquilla save 60-80% on token costs?
OpenSquilla uses multiple coordinated strategies: ML classifier routing that evaluates request complexity using message length, code blocks, keywords, and embeddings to route simple tasks to cheaper models; reasoning depth tiers that disable deep thought billing for simple queries; adaptive prompts that tune thinking depth based on task; and on-demand skills that only load needed capabilities into context.
What is the four-tier memory system?
The four-tier cognitive memory architecture mirrors human cognition: Working memory for current task context, Episodic memory for experience and causality, Semantic memory for facts and rules, and Raw memory as audit and retraining base. It includes vector-semantic plus BM25 keyword hybrid search, hot memory promotion, temporal decay, and automated Memory Dream consolidation every 24 hours.
How does the security sandbox work?
OpenSquilla uses syscall-level isolation through Bubblewrap on Linux and Seatbelt on macOS, not Docker. It has a three-tier policy system: Standard runs directly, Strict requires sandbox approval, Locked enforces human review. Additional protections include denial ledger (auto-pause after 3 rejections), stale output protection (purge rejected results), and prompt injection defense via XML-escaping skill metadata and tool results.
What platforms does OpenSquilla support?
OpenSquilla runs on Windows, macOS, and Linux. It requires Python 3.12+. The portable Windows version includes bundled CPython so no separate Python installation is needed. Sandbox isolation uses Bubblewrap on Linux and Seatbelt on macOS.
How do I install OpenSquilla?
The recommended installation uses uv tool install: first install uv via curl, then run 'uv tool install --python 3.12 opensquilla[recommended] @ [release wheel URL]'. After installation, run 'opensquilla onboard' for interactive setup wizard, then 'opensquilla gateway run' to start the ASGI server. Alternatively, download the portable Windows ZIP from GitHub releases.
What communication channels are built-in?
OpenSquilla includes 10+ built-in channels: Terminal, Web UI, Slack, Discord, Telegram, MS Teams, Matrix, Lark, DingTalk, WeCom, and QQ. You configure one agent and serve users across all these channels simultaneously via the ASGI RPC gateway.
Can I use OpenSquilla completely offline?
Yes, OpenSquilla supports fully offline/on-prem deployment. Data never leaves your network. It includes bundled ONNX local inference for embeddings that runs on CPU, and built-in web search can be configured. ML routing runs locally without external dependencies once installed.
How do I migrate from OpenClaw or Hermes Agent?
OpenSquilla offers one-click migration from both OpenClaw and Hermes Agent frameworks. The microkernel architecture allows easy plugin implementation, and the framework is designed with compatibility in mind for users switching from these peer open-source agent frameworks.
Is there a free tier or trial?
OpenSquilla itself is fully open source under Apache 2.0, free to pull from GitHub and self-host anytime. The website offers a limited-time free token credit trial (zero credit card required, 10-second signup) so users can verify the 60-80% cost savings with zero risk. Free tokens are limited quantity, first-come-first-served. Users can also use the open-source version with their own tokens.