OpenSquilla Unveils Open-Source AI Agent That Cuts Token Costs

OpenSquilla has unveiled a v0.1.0 self-hostable open-source AI agent runtime. The core goal is to lower the cost ceiling of long-running tasks by reducing tokens that agents waste unnecessarily.

In local testing, three prompts processed a total of 279,762 tokens, with a session cost of $0.0094. Of these, 222,848 tokens were handled via cache, reusing about 80% of input tokens. The routing classifier logged the gate decision for each request, and while the sandbox runs as a no-op on Windows, it provides syscall-level isolation in Linux production.

Cost reduction is implemented through multiple layers of routing.

• An ML classifier judges request complexity using message length, code blocks, keyword patterns, and embedding features together.
• Simple requests are routed to cheaper models, and deep reasoning is turned off for lightweight tasks.
• Skills are loaded on-demand rather than being placed in the full context.
• Based on the team's own benchmarks, this enables 60-80% savings in token spending compared to a fixed single-model configuration.
• Quota hooks and per-call cost tracking detect and block overspending.

Memory is divided into a 4-tier cognitive architecture. Working memory handles the current task, Episodic memory handles experiences and causal relationships across sessions, Semantic memory handles persistent facts and rules, and Raw memory handles originals for auditing and retraining.

Retrieval runs vector-semantic search and BM25 full-text search in parallel, with embeddings processed locally via bundled ONNX inference. Frequently referenced items are promoted to hot memory, older memories are diluted through temporal decay, and every 24 hours, Memory Dream Consolidation reorganizes scattered memories into denser knowledge.

Security uses syscall-level isolation instead of a Docker wrapper. Policies come in three tiers—standard, strict, and locked—with Bubblewrap on Linux and Seatbelt on macOS separating filesystem access. After three consecutive denials, a denial ledger halts the agent, and to prevent prompt injection, skill metadata and tool results are XML-escaped.

The architecture is a microkernel centered on a roughly 100-line core orchestrator. Providers, memory backends, channel adapters, and tool integrations all attach as plugins, and plugins can be written as 5-line duck-typed classes without a base class or manifest. It supports more than 10 channels including Slack, Discord, Telegram, MS Teams, and Matrix, and the release is based on Apache-2.0 and Python 3.12+, available for self-hosting on GitHub. Alongside the launch, the team is also running a 10M Token Bill Challenge.