OpenSquilla — Token-Efficient AI Agent

Overview

OpenSquilla is a token-efficient, microkernel AI agent — same budget, more capability, better results. It combines smart routing, persistent memory, a secure sandbox, built-in web search, and local embeddings under a single model loop. Every entry point — Web UI, CLI, and chat channels — runs through a shared TurnRunner, and a pluggable provider layer lets it speak to OpenRouter, OpenAI, Anthropic, Ollama, DeepSeek, Gemini, Qwen/DashScope, and roughly twenty other LLM providers without changes to your code or config schema.

Quick start

The fastest path to a running OpenSquilla on your local machine.

Choose one path and stay on it:

Goal	Use this path	Commands use
Run OpenSquilla as a local app	Install	opensquilla ...
Modify or debug OpenSquilla source	Develop from source	uv run opensquilla ...

Both paths can start from git clone. In the install path, the clone is only the source package the installer reads from. In the development path, the clone is the live workspace.

1. Install prerequisites: git-lfs and uv.

2. Clone with LFS assets:

   git lfs install
   git clone https://github.com/opensquilla/opensquilla.git
   cd opensquilla
   git lfs pull --include="src/opensquilla/squilla_router/models/**"

3. Install with the recommended profile. This creates a user-local opensquilla command. The checkout-local .venv, if any, is not used.

   macOS / Linux:

   bash install.sh

   Windows PowerShell:

   pwsh -ExecutionPolicy Bypass -File install.ps1

   Optional channel adapters are installed only when requested. For example, add Feishu websocket support with -Extras feishu on Windows or OPENSQUILLA_INSTALL_EXTRAS=feishu on macOS/Linux.

   Only set OPENSQUILLA_INSTALL_PROFILE=core if you intentionally want to skip the bundled router.

4. Configure. Use the installed opensquilla command below. Do not prefix these commands with uv run unless you chose Develop from source.

   opensquilla onboard

5. Run the gateway:

   opensquilla gateway run

Open the Web UI at http://127.0.0.1:18790/control/.

Advanced usage

A complete tutorial that covers every step in Quick start plus the options Quick start glosses over. Sections marked (optional) can be skipped depending on your environment; everything else is required for a working install.

Prerequisites

• Python 3.12+ — required for source and uv installs. (optional) for portable-zip users, since the release zip already bundles its own CPython.
• Git and Git LFS — required. The bundled SquillaRouter assets are stored as LFS pointers; without git-lfs the recommended profile fails with version https://git-lfs.github.com/spec/v1 pointer files in place of the model bytes. Install once: https://git-lfs.com/.
• uv or pip ≥ 23 — required. The installer scripts prefer uv tool install and fall back to pip --user. Install uv once: https://docs.astral.sh/uv/.

Clone the repo

git lfs install
git clone https://github.com/opensquilla/opensquilla.git
cd opensquilla
git lfs pull --include="src/opensquilla/squilla_router/models/**"

Install

Use this path when you want to run OpenSquilla, not edit its source. The clone is only the package source for the installer. After install, use opensquilla ...; do not use uv run.

The scripts install .[recommended] by default. recommended is the normal runtime profile: router, memory, and local model dependencies. Messaging channel adapters are opt-in extras. Most users do not need every chat platform SDK.

macOS / Linux:

bash install.sh

Windows PowerShell:

pwsh -ExecutionPolicy Bypass -File install.ps1

Install channel extras into the same user-local command:

Windows PowerShell:

pwsh -ExecutionPolicy Bypass -File install.ps1 -Extras feishu

macOS/Linux:

OPENSQUILLA_INSTALL_EXTRAS=feishu bash install.sh

Supported extras include feishu, telegram, dingtalk, wecom, qq, msteams, matrix, matrix-e2e, and document-extras.

The scripts prefer uv tool install and fall back to python -m pip install --user. The installed command uses its own Python environment; it is separate from a checkout-local .venv.

Useful install options:

$env:OPENSQUILLA_INSTALL_PROFILE="core"          # minimal runtime
$env:OPENSQUILLA_INSTALL_DRY_RUN="1"             # print the plan only

OPENSQUILLA_INSTALL_PROFILE=core bash install.sh
OPENSQUILLA_INSTALL_DRY_RUN=1 bash install.sh

To check which command your shell will run:

where.exe opensquilla

command -v opensquilla

After reinstalling from a local checkout, restart the gateway process so it loads the updated package.

Develop from source

Use this path only when you want to modify, test, or debug the current checkout. uv sync creates the checkout-local .venv, and uv run executes against the live source tree.

uv sync --extra recommended
uv run opensquilla --help

Install extras into the same environment you run:

uv sync --extra recommended --extra feishu
uv run opensquilla channels status feishu --json

In this mode, prefix every command below with uv run. Do not debug a development checkout through a user-local opensquilla command; that command runs in a different Python environment.

First-run config

opensquilla onboard --if-needed is the recommended post-install entrypoint for first-run setup and automation. It writes the active config file, skips when an LLM provider is already configured, and keeps provider secrets in environment variables when you pass --api-key-env. The router defaults to recommended, which enables SquillaRouter for supported provider profiles. Pass --router disabled only if you intentionally want direct single-model routing, or --router openrouter-mix to keep the built-in OpenRouter mixed model routes. Useful invocations:

opensquilla onboard                # full interactive wizard
opensquilla onboard --if-needed    # idempotent: skip if already configured
opensquilla onboard --minimal      # provider only, skip channels/search

In SSH, CI, or any environment without a TTY the interactive flow exits with code 2. Use the non-interactive form — keep the secret in the environment and pass its name, not its value, to onboard:

export OPENROUTER_API_KEY="sk-..."
opensquilla onboard \
  --provider openrouter \
  --api-key-env OPENROUTER_API_KEY

(optional) Re-configure one section later without redoing the whole wizard:

opensquilla configure provider --provider openai --model gpt-4o
opensquilla configure router --router recommended
opensquilla configure search   --search-provider brave
opensquilla configure channels                # interactive section

Sections: provider, router, channels, search, image-generation, memory-embedding. The Web UI also exposes a setup flow at /control/setup for provider, router tiers, optional channels, and extras. Later CLI edits should use `opensquilla configure

` rather than provider-specific aliases.

Messaging channel saves are config changes, not runtime connectivity proof. Restart the gateway process after channel edits, then verify the live adapter state:

opensquilla gateway restart
opensquilla channels status <name> --json

Treat a channel as connected only when the status payload reports enabled=true, configured=true, and connected=true. Feishu defaults to websocket mode and does not need a public URL in that mode; Feishu webhook mode, Slack, WeCom, and Microsoft Teams require a public provider-reachable URL.

Config load order: OPENSQUILLA_GATEWAY_CONFIG_PATH → ./opensquilla.toml → ~/.opensquilla/config.toml → built-in defaults. Onboarding writes the file at the path the runtime would read; environment values for individual secrets always win over file values.

Run

opensquilla gateway run                   # foreground, 127.0.0.1:18790
opensquilla gateway start --json          # background + health wait
opensquilla chat                          # interactive REPL
opensquilla agent -m "your prompt"        # one-shot, automation-friendly

Open the Web UI at http://127.0.0.1:18790/control/ and check health with curl http://127.0.0.1:18790/health.

Public network binding — (optional)

To make the Web UI reachable from another machine, bind the gateway to all interfaces and use the host's public IP address:

opensquilla gateway run --listen 0.0.0.0 --port 18790
# or, for a background process:
opensquilla gateway start --listen 0.0.0.0 --port 18790 --json

Then open http://<public-ip>:18790/control/ and verify the public health endpoint with:

curl http://<public-ip>:18790/health

If another gateway is already bound to 18790, stop it first or choose a different --port. Public access also requires the host firewall or cloud security group to allow inbound TCP traffic on that port. Do not expose the gateway publicly with [auth] mode = "none"; configure token or password auth before binding to 0.0.0.0.

Docker and portable paths — (optional)

./start.sh (or start.ps1 on Windows) wraps docker compose up -d and tails the gateway logs — convenient if you do not want a Python toolchain on the host. Release zips that bundle a CPython runtime are produced by the Wheelhouse Zip Release workflow; portable users extract the zip and run its bundled launcher without a system Python install.

Further tuning

Provider-specific config, tier profiles, sandbox tuning, image generation, and concurrency settings are managed through opensquilla onboard, opensquilla config, and opensquilla.toml.example.

Benchmark Results

PinchBench 1.2.1 average results across 25 tasks:

Agent	Base Model	Avg. score	Total input tokens	Total output tokens	Total cost
OpenSquilla	Model router (Opus4.7, GLM5.1, DS4 Flash)	0.9251	1,721,328	61,475	$0.688
OpenClaw	Claude Opus 4.7	0.9255	3,066,243	50,890	$6.233

Key Features

• Token-efficient routing — local SquillaRouter (LightGBM + ONNX BGE classifier, recommended extra) routes each turn across four tiers (T0–T3). Hybrid features (length, language, code blocks, keywords + semantic embeddings) pick the cheapest model that can handle the turn; classification runs on-device, so your prompt never leaves the machine to make the decision.
• Adaptive reasoning and prompts — reasoning-token billing only kicks in when the turn needs deep thought, and the system prompt scales with task complexity (lightweight for trivial turns, full instructions for complex ones). No paying reasoning tokens for "hello".
• On-demand skills — built-in MCP client plus 16 bundled skills (coding agents, GitHub, cron, deep research, pptx/docx/xlsx/pdf toolkits, summarization, tmux, weather, and more); only the skills needed for the current task are loaded into context, avoiding steady-state token waste.
• Four-tier cognitive memory — working (current task) → episodic (experience and causality) → semantic (facts and rules) → raw (audit and retraining base), mirroring human cognition.
• Hybrid memory search + local embeddings — Markdown source-of-truth memory with FTS keyword search alongside sqlite-vec semantic recall. Bundled ONNX inference runs on CPU so embeddings stay on your machine; optionally swap to OpenAI- or Ollama-hosted embeddings.
• Adaptive recall and consolidation — frequently used memories auto-promote and dated ones decay exponentially (with an "evergreen" opt-out); periodic Dream consolidation merges scattered episodic traces into structured knowledge, mirroring sleep consolidation, with bounded prompt-injection budgets throughout.
• Layered security sandbox — three policy tiers (Standard / Strict / Locked) on a permission-tier matrix, with Bubblewrap on Linux executing code in isolated environments (the macOS Seatbelt backend currently renders SBPL profiles only; process execution is pending). A denial ledger auto-pauses autonomous execution after repeated sandbox denials, rejected outputs are purged via intent + stale-output caches so the agent can't recover them through a side channel, and all skill metadata and tool results are XML-escaped to close common prompt-injection vectors.
• Unified gateway across all entry points — Starlette ASGI server on 127.0.0.1:18790 with WebSocket RPC and an embedded control console (/control/). Web UI, CLI, and first-class adapters for Terminal, WebSocket, Slack, Telegram, Discord, Feishu, DingTalk, WeCom, MS Teams, Matrix, and QQ all converge on a shared TurnRunner for consistent tool dispatch, retry, and decision logging.
• 20+ LLM providers — OpenRouter, OpenAI, Anthropic, Ollama, DeepSeek, Gemini, DashScope/Qwen, Moonshot, Mistral, Groq, Zhipu, SiliconFlow, Volcengine, BytePlus, MiniMax, vLLM, LM Studio, OVMS, and more, with a primary-plus-fallback selector.
• Durable sessions, agents, and scheduling — SQLite-backed session, transcript, and replay storage with per-agent workspaces and a reset/flush contract that proves persistence before destructive rewrites; SchedulerEngine with an in-tree CronExpression parser plus stagger, reaper, and heartbeat services exposed via the opensquilla cron CLI.

Credits

OpenSquilla is a token-efficient AI Agent inspired by OpenClaw. Bundled third-party content is fully attributed in THIRD_PARTY_NOTICES.md.

Contributing

OpenSquilla is an open-source project and we welcome contributions of every kind — bug reports, feature ideas, documentation, new provider or channel adapters, skills, and core runtime work. Open an issue or a pull request on GitHub to get involved.