TokenRhythm/opensquilla

OpenSquilla — Token-Efficient AI Agent with same budget, higher intelligence density

OpenSquilla – a token‑efficient micro‑kernel AI agent

What it is – OpenSquilla is a locally‑run AI‑agent framework that routes each user turn to the cheapest LLM that can handle the request. It combines a lightweight on‑device model router (SquillaRouter), persistent memory, a sandboxed tool layer, built‑in web search and on‑device embeddings, all inside a single “turn loop”. The same loop powers a desktop GUI, a CLI, and chat‑channel integrations, so behaviour is identical no matter how you talk to the agent.

Why it matters – By automatically picking the most cost‑effective model for each turn, OpenSquilla reduces token spend while keeping capability. The router can also ensemble multiple models, a technique the authors claim outperforms the “Fable 5” benchmark.

Key components

Component Role
SquillaRouter On‑device router that decides which LLM to call (supports ONNX Runtime, LightGBM, etc.)
Provider layer Uniform API for >20 LLM providers (TokenRhythm, OpenRouter, OpenAI, Anthropic, Ollama, Gemini, Qwen/DashScope, …)
Persistent memory Stores long‑term context across sessions
Layered sandbox Isolates tool execution and retries, ensuring safe tool use
Built‑in web search & embeddings Enables retrieval‑augmented generation without external services
Control console Vue‑based web UI packaged in an Electron shell for desktop use

Installation paths (choose one)

  1. Desktop installers – pre‑built .dmg (mac Apple Silicon) or .exe (Windows) that include the Vue console and all native runtimes. No Git or Node needed.
  2. Quick terminal install – recommended for all OSes. Uses the uv Python‑package manager to install a wheel from the GitHub release:
    uv tool install --python 3.12 "opensquilla[recommended] @ https://github.com/TokenRhythm/opensquilla/releases/download/v0.5.4/opensquilla-0.5.4-py3-none-any.whl"
opensquilla onboard
opensquilla gateway run
  3. Install from source – clone the repo (Git LFS pulls model assets), run the provided scripts/install_source.sh/.ps1 which builds the Vue UI, installs dependencies, and registers the opensquilla command.
  4. Develop from source – for contributors. Uses uv sync to create a local editable environment, then uv run opensquilla ….

Note – The router needs native runtimes (ONNX Runtime on Windows, libomp on macOS). The desktop installers bundle them; the terminal install may require you to install the Visual C++ redistributable (Windows) or brew install libomp (macOS).

Typical workflow

# Set up the agent (once)
opensquilla onboard          # creates config, generates install_id, runs telemetry opt‑out prompts

# Start the gateway (the long‑running process that handles turns)
opensquilla gateway run

# Interact via CLI, Web UI, or a connected channel (Slack, Discord, Matrix, etc.)
opensquilla chat "Summarize the latest news about AI safety"

The gateway will:

  1. Receive the user turn.
  2. Ask SquillaRouter which provider can answer within the token budget.
  3. Call the selected LLM, optionally augmenting the prompt with retrieved web results or embeddings.
  4. Store the turn in persistent memory and log the routing decision.
  5. Return the response to the originating interface.

Supported integrations (out‑of‑the‑box)

  • Feishu, Telegram, DingTalk, QQ, WeCom, Slack, Discord
  • Optional extras: Matrix (with optional end‑to‑end encryption), PDF generation via WeasyPrint, etc.

Privacy & telemetry

OpenSquilla sends pseudonymous telemetry on first start and once per day: install ID (hashed MAC/IP), version, OS, token‑usage aggregates, and install method. No content, API keys, or personal identifiers are transmitted. Telemetry and automatic update checks can be disabled via:

[privacy]
disable_network_observability = true

or the legacy env vars OPENSQUILLA_TELEMETRY_DISABLED / OPENSQUILLA_UPDATE_CHECK_DISABLED.

Licensing & community

  • License: Apache 2.0 (permissive, allows commercial use).
  • Documentation: Product guide, CLI reference, and a technical report (arXiv 2607.11399) are bundled in the repo.
  • Release status: 0.5.4 (stable). Ongoing development is tracked on the main branch; CI badges show automated testing.
  • Support: Issues can be opened on GitHub; the website opensquilla.ai provides links to the report, docs, and contact information.

Who might benefit?

  • Developers who want a locally‑run, cost‑aware LLM agent without wiring together separate routers.
  • Product teams needing a single binary that works on Windows/macOS/Linux and offers a ready‑made desktop UI.
  • Teams with strict data‑privacy requirements, because the core routing and memory stay on‑device.
  • Researchers interested in “agentic routing” and the data‑flywheel concept described in the accompanying technical report.

Quick start checklist

  1. Install uv (or have Python 3.12+ and Node 22.12+ if building from source).
  2. Choose an install path (desktop installer or uv tool install).
  3. Run opensquilla onboard to generate config.
  4. Start the gateway with opensquilla gateway run.
  5. Interact via opensquilla chat or open the web UI at http://localhost:….

Bottom line: OpenSquilla is a real, production‑grade AI‑agent platform that focuses on token‑efficient routing, unified interfaces, and on‑device privacy. It is well‑documented, cross‑platform, and released under a permissive open‑source license.

Related

  • Project
    espeak-ng/espeak-ng

    A compact, open-source text-to-speech synthesizer using formant synthesis to support over 100 languages with a very small memory footprint.

  • Project
    akitaonrails/ai-memory

    ai‑memory is a Rust‑based server that records every observation a coding‑assistant makes (prompts, tool calls, session boundaries) as plain markdown files in a git‑backed wiki. The server can be run locally, in Docker/Podman, or on a homelab box and is designed to be shared across many agents (Claude Code, Codex, Cursor, Gemini CLI, etc.), machines and team members. Hooks posted by agents are consolidated into readable wiki pages at session end; the next session receives a concise handoff so it can continue exactly where the previous one stopped. Storage is human‑readable markdown plus a SQLite index (FTS5, entity and optional vector search). By default it needs no LLM or embedding keys; optional Anthropic/OpenAI integrations add LLM‑generated summaries and semantic search. Multi‑user authentication, audit logs and per‑person attribution are built in. Install via Arch AUR, Docker wrapper, or native binary, start the server, register a client and its hooks, and the system works transparently.

  • Project
    TraderAlice/OpenAlice

    OpenAlice is a locally‑run AI orchestrator for individual traders. It lets you plug in LLM agents (Claude, Codex, Pi, etc.) and connect them to market data, research workspaces (Git‑backed folders), and broker accounts. Agents can fetch data, run quantitative analyses, generate reports, and even draft trade orders, which you review and commit via a Git‑style interface. The product ships as a desktop app (macOS/Windows) with CLI/Docker options, and includes scheduling, issue tracking, and a central Inbox for follow‑up. Trading execution is marked as beta and intended for paper or test accounts first.

  • Project
    trailofbits/skills

    Trail of Bits Skills is a Claude Code/Codex plugin marketplace that adds dozens of security‑focused “skills” (e.g., static analysis, smart‑contract auditing, YARA rule authoring, mutation testing). Install it with a single `/plugin marketplace add trailofbits/skills` command, then invoke individual plugins from the LLM to run concrete analyses and get structured results.