TokenRhythm/opensquilla

Same budget, more capability, better results.
A microkernel AI agent for your CLI, Web UI, and chat channels.

English · 中文 · 日本語 · Français · Deutsch · Español

• 📢 2026-08-22 — The English version of our technical report is now on aiXiv: aixiv.260822.000001, and the Chinese version is on ChinaXiv: 202608.00176. See Citation for how to cite OpenSquilla.

• 📢 2026-07-14 — Our technical report Agentic Routing: The Harness-Native Data Flywheel is now on arXiv. It shows how the harness-native router turns everyday agent traffic into a self-improving data flywheel, and how multi-model ensemble routing surpasses Fable 5.

OpenSquilla is a token-efficient, microkernel AI agent. A local model router sends each turn to the cheapest model that can handle it, while persistent memory, a layered sandbox, built-in web search, and on-device embeddings round out a single shared turn loop.

Every entry point — Web UI, CLI, and chat channels — runs through that same loop, so tool dispatch, retries, and decision logging behave identically everywhere. A pluggable provider layer speaks to TokenRhythm, OpenRouter, OpenAI, Anthropic, Ollama, DeepSeek, Gemini, Qwen/DashScope, and 20+ other LLM providers with no change to your code or config schema.

OpenSquilla 0.5.6 is the current stable release.

For task-oriented product documentation, start with the OpenSquilla Product Guide or the documentation index.

OpenSquilla runs on Windows, macOS, and Linux. Pick the path that matches your use case.

Desktop installers and Quick terminal install give you a prebuilt release — no Git required. The other two — Install from source and Develop from source — build from a Git checkout (git clone + Git LFS), including the Vue control console. Release wheels and Desktop installers already contain that console, so their users do not need Node.js or npm.

Release install commands use published GitHub release assets. Python wheel installs use versioned wheel filenames because installers validate the version embedded in the wheel filename.

For 0.5.6 desktop use, prefer the packaged desktop installers from the GitHub Release: OpenSquilla-0.5.6-mac-arm64.dmg on macOS and OpenSquilla-0.5.6-win-x64.exe on Windows.

The default recommended profile installs SquillaRouter — OpenSquilla's on-device model router — and its model assets; OPENSQUILLA_INSTALL_PROFILE=core omits those dependencies. The separate --router disabled onboarding flag keeps the dependencies installed but turns the router off at runtime.

On Windows, SquillaRouter's bundled ONNX runtime also needs the Visual C++ runtime. The from-source PowerShell installer installs it automatically via winget; the Quick terminal install (uv tool install) path does not — if startup logs a DLL load failed error, install it manually (see Troubleshooting). OpenSquilla keeps running with direct single-model routing until it is installed.

On macOS terminal installs, SquillaRouter's LightGBM runtime may also need the system OpenMP library. The desktop app bundles the runtime it needs, but Quick terminal install does not install Homebrew/system libraries. If startup logs Library not loaded: @rpath/libomp.dylib, run brew install libomp, then restart the gateway. OpenSquilla keeps running with direct single-model routing until it is installed.

Install links: Git · Git LFS · Node.js · uv.

The 0.5.6 desktop installers package the Vue control console and gateway runtime in an Electron shell.

• macOS Apple Silicon: https://github.com/TokenRhythm/opensquilla/releases/download/v0.5.6/OpenSquilla-0.5.6-mac-arm64.dmg
• Windows x64: https://github.com/TokenRhythm/opensquilla/releases/download/v0.5.6/OpenSquilla-0.5.6-win-x64.exe

For faster Mainland China downloads, use the OSS direct-download aliases:

• macOS Apple Silicon: https://opensquilla-releases.oss-cn-beijing.aliyuncs.com/releases/latest/OpenSquilla-mac-arm64.dmg
• Windows x64: https://opensquilla-releases.oss-cn-beijing.aliyuncs.com/releases/latest/OpenSquilla-win-x64.exe

These fixed links advance only after a newer eligible release passes mirror verification. Use the versioned GitHub Release links above when you need a specific release.

Quit any running OpenSquilla desktop app before upgrading. On macOS, drag the app from the DMG into Applications for installation or updates, eject the DMG, then open the Applications copy. The existing Desktop profile in the platform application-data directory is reused. A terminal installation's ~/.opensquilla is a separate profile; transfer it explicitly from Settings only if needed.

When upgrading the Windows Desktop from RC3 to RC4 or later, run the new installer directly over the existing installation. Do not uninstall RC3 first: its uninstaller may remove Desktop user data. Back up %APPDATA%\OpenSquilla before upgrading. RC4 and later installers preserve profile data during a normal uninstall.

Code signing policy: docs/code-signing-policy.md.

The recommended path on Windows, macOS, and Linux. uv installs OpenSquilla into its own isolated environment and manages its own Python — no system Python required. This path installs published releases only; for main, development branches, or local checkouts use Install from source.

1. Install uv — skip if uv --version already works.

Linux / macOS:

curl -LsSf https://astral.sh/uv/install.sh | sh
. "$HOME/.local/bin/env"

Windows PowerShell:

powershell -c "irm https://astral.sh/uv/install.ps1 | iex"
$env:Path = "$env:USERPROFILE\.local\bin;" + $env:Path

2. Install OpenSquilla — the same command on every platform.

uv tool install --python 3.12 "opensquilla[recommended] @ https://github.com/TokenRhythm/opensquilla/releases/download/v0.5.6/opensquilla-0.5.6-py3-none-any.whl"

This installs the OpenSquilla wheel from the release URL, then lets uv download the dependencies declared by the selected extras. The default recommended extra includes SquillaRouter runtime dependencies such as ONNX Runtime, LightGBM, NumPy, and tokenizers, so a first install needs network access unless those wheels are already cached. uv does not install system native runtimes such as macOS libomp or the Windows Visual C++ Redistributable; see Troubleshooting if the router runtime reports a native-library load error.

3. Configure and run.

opensquilla onboard
opensquilla gateway run

For a fully pinned install, use the versioned wheel URL: https://github.com/TokenRhythm/opensquilla/releases/download/v0.5.6/opensquilla-0.5.6-py3-none-any.whl.

Use this path to run OpenSquilla from a checkout without editing it. The clone is only the package source for the installer; after install, use the opensquilla command — do not run uv run. Choose Develop from source instead if you intend to modify the code.

1. Clone with LFS assets

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

2. Run the installer

   macOS / Linux

   bash scripts/install_source.sh

   Windows PowerShell

   powershell -ExecutionPolicy Bypass -File ./scripts/install_source.ps1

   The script installs .[recommended] (SquillaRouter + memory + local models) into a dedicated user environment via uv tool install. Before the Python install, it runs npm ci and npm run build in opensquilla-webui. Every source reinstall recreates the locked node_modules tree and rebuilds the console; the first run normally has the largest dependency download, while a warm npm cache reduces later network use but not all build time or disk writes. It then installs the built console with the Python package, falling back to python -m pip install --user when uv is unavailable. If opensquilla is not on PATH after install (common on a fresh host where ~/.local/bin is not yet on PATH), run uv tool update-shell and open a new terminal; see Troubleshooting for details.

   Direct pip install ., uv tool install ., and VCS URL installs are low-level source-build paths, not substitutes for this installer. A local checkout works only after its Web UI has been built; a VCS URL checkout has no generated artifact and is intentionally rejected. Use this source installer or an official release wheel instead.

3. (optional) Install advanced extras. Most channels — Feishu, Telegram, DingTalk, QQ, WeCom, Slack, and Discord — work from the base install. The opt-in extras are:

   • matrix — Matrix channel (pulls in matrix-nio)
   • matrix-e2e — Matrix channel with end-to-end encryption (requires libolm)
   • document-extras — PDF generation via WeasyPrint

   OPENSQUILLA_INSTALL_EXTRAS=matrix bash scripts/install_source.sh        # macOS / Linux

   powershell -ExecutionPolicy Bypass -File ./scripts/install_source.ps1 -Extras matrix   # Windows

4. Configure and run — see Configuration.

winget install --id Git.Git -e
winget install --id GitHub.GitLFS -e
winget install --id OpenJS.NodeJS.LTS -e
powershell -ExecutionPolicy Bypass -c "irm https://astral.sh/uv/install.ps1 | iex"
git lfs install

brew install git git-lfs node uv
git lfs install

sudo apt update && sudo apt install -y git git-lfs curl
curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
sudo apt install -y nodejs
curl -LsSf https://astral.sh/uv/install.sh | sh
git lfs install

OPENSQUILLA_INSTALL_PROFILE=core   bash scripts/install_source.sh   # minimal runtime, no SquillaRouter
OPENSQUILLA_INSTALL_DRY_RUN=1      bash scripts/install_source.sh   # print the plan only

Use this path when you are working on OpenSquilla's source code: making changes, running tests, or debugging behavior against this checkout. It is not the normal install path. Unlike Install from source, this path requires uv: uv sync creates a repository-local .venv, and uv run executes commands against the files in this checkout.

cd opensquilla-webui
npm ci
npm run build
cd ..
uv sync --extra recommended --extra dev
uv run opensquilla --help

Run npm run build again after changing Web UI sources. Standard wheel builds fail closed when the generated console is missing or stale; editable uv sync installs remain available for backend-only work.

The recommended extra includes SquillaRouter for development too; the dev extra installs the test, lint, and typecheck tools. Install additional extras into the same environment you run:

uv sync --extra recommended --extra dev --extra matrix
uv run opensquilla channels status matrix --json

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

Remove OpenSquilla with opensquilla uninstall. It keeps your data by default and removes only the program:

opensquilla uninstall --dry-run   # preview what would be removed and kept
opensquilla uninstall             # remove the program, keep your data

To delete data too, opt in explicitly:

opensquilla uninstall --purge-state    # sessions, logs, cache, scheduler, memory
opensquilla uninstall --purge-config   # config.toml and secrets (.env)
opensquilla uninstall --purge-all      # everything (asks you to type a confirmation)

The running gateway is drained and stopped first, deletion stays inside the OpenSquilla home, and Docker/desktop installs get guided removal steps instead. Desktop or OS app removal remains platform-specific; the CLI guidance does not remove a desktop app bundle. See docs/cli.md for the full reference.

OpenSquilla uses the existing Network reporting switch for V1 statistics and both V2 statistics streams. Reporting is enabled by default and can be turned off in Privacy settings, without separate onboarding choices or consent popups:

• Reliability diagnostics records bounded operation results for app and Gateway startup, crashes, turns, tools, file parsing, updates, and session performance.
• Product and growth analytics records client launches, product activity, and one-time acquisition, onboarding, app-readiness, registration, and first-successful-turn milestones. Existing installations do not become new-user cohorts just by enabling reporting.

The streams retain separate session/journey identifiers, durable queues, upload endpoints, and retention policies. Reliability events go to /v1/reliability/events; growth events go to /v1/growth/events. Retries reuse event_id for server-side deduplication, and growth events are not sampled.

Application events also carry a one-way, application-specific device_id derived locally from the OS machine identifier when available. Daily/monthly activity and feature device counts deduplicate this token across profiles and client surfaces. Legacy events without it are excluded from device counts; operation totals and success rates still count actual operations.

V2 statistics never include prompts, responses, file names, file paths, file contents, tool arguments, task parameters, provider configuration, raw account IDs, order data, MAC addresses, IP addresses, or raw OS machine identifiers. Complete crash stacks stay local unless the user explicitly prepares and shares a support bundle.

V1 installation/version reporting at /v1/install and daily conversation/token totals at /v1/usage run alongside V2 after Gateway readiness. Only completed UTC days are uploaded; pending days retry hourly. Existing installation state is retained. Daily event IDs use a persistent random identity per aggregate database to separate profiles while keeping retries stable. V1 retains its pseudonymous installation ID derived locally from MAC/IP data, with a persisted random fallback; raw MAC/IP values are not uploaded. The X-OpenSquilla-Install-Id provider header remains retired, and V2 keeps its independent identities.

To force all non-user-initiated network observability off before startup:

OPENSQUILLA_PRIVACY_DISABLE_NETWORK_OBSERVABILITY=true

or set:

[privacy]
disable_network_observability = true

This is a hard veto over V1 and both V2 statistics streams, passive update checks, and automatic desktop update checks. Disabling Network reporting pauses pending uploads and stops collection without deleting local statistics state. Previously saved per-scope declines are migrated to the unified switch being off; users can then change that one setting. CI, test, and DO_NOT_TRACK environments also fail closed for statistics uploads. Other user-initiated actions may still contact configured providers, search services, channels, or release hosts. Explicit update-availability checks remain disabled while the unified or legacy update opt-out controls are active.

Legacy opt-out environment variables remain honored:

OPENSQUILLA_TELEMETRY_DISABLED=true
OPENSQUILLA_UPDATE_CHECK_DISABLED=true

The legacy statistics variable disables V1 and V2 reporting. The legacy update variable also suppresses V1 uploads for compatibility, but does not disable V2. See PRIVACY.md for the complete data, reporting, deletion, update, and external-producer rules.

opensquilla onboard is the interactive first-run wizard. It writes the active config file and keeps provider secrets in environment variables when you pass --api-key-env. The router defaults to recommended (SquillaRouter on supported providers); pass --router disabled for direct single-model routing.

opensquilla onboard                # full interactive wizard
opensquilla onboard --if-needed    # idempotent: safe for scripts and re-installs
opensquilla onboard --minimal      # provider only; skip channels and search
opensquilla onboard status         # inspect every setup section without writing

In SSH, CI, or any environment without a TTY, use the non-interactive form — keep the secret in the environment and pass its name, not its value:

Linux / macOS

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

Windows PowerShell

$env:OPENROUTER_API_KEY="sk-..."
opensquilla onboard --provider openrouter --api-key-env OPENROUTER_API_KEY

OpenRouter is only an example — substitute any supported provider and its API-key variable.

Re-configure one section later without redoing the whole wizard (these examples assume the relevant API key is already in the environment):

opensquilla configure provider --provider openai --model gpt-4o --api-key-env OPENAI_API_KEY
opensquilla configure router --router recommended
opensquilla configure search   --search-provider duckduckgo
opensquilla configure search   --search-provider exa --api-key-env EXA_API_KEY
opensquilla configure channels

Sections: provider, router, channels, search, image-generation, memory-embedding. The Web UI exposes the same catalog and status model at /control/setup: Provider and Router are the fast path, while Channels, Search, Image generation, and Memory embedding sit in the Capability Center and can be configured later. Empty channels are treated as an opt-out, not a failed setup.

Config load order: OPENSQUILLA_GATEWAY_CONFIG_PATH → ./opensquilla.toml → ~/.opensquilla/config.toml → built-in defaults. Environment values for individual secrets always win over file values.

If you already have state under ~/.openclaw or ~/.hermes, run a dry run first to inspect the migration report, then apply it explicitly:

opensquilla migrate openclaw --json
opensquilla migrate openclaw --apply

opensquilla migrate hermes --json
opensquilla migrate hermes --apply

Use opensquilla migrate --source openclaw,hermes --apply to import both default homes. Add --migrate-secrets only after reviewing the dry-run report. See MIGRATION.md for custom paths and conflict handling.

opensquilla gateway run                # foreground, 127.0.0.1:18791
opensquilla gateway start --json       # background + health wait
opensquilla chat                       # interactive REPL
opensquilla agent -m "your prompt"     # one-shot, automation-friendly

For subprocess progress, add --event-stream-stderr. The final result keeps its existing stdout format, while stderr receives incrementally flushed, privacy-bounded v1 JSONL events. Consumers must drain stderr continuously and only treat objects with "_event": true as events because ordinary diagnostics can share the stream. See docs/cli.md for the versioned schema and compatibility contract.

Parallel agent subprocesses must use separate profile homes and persistent state roots. Set both variables for every child:

OPENSQUILLA_STATE_DIR=/tmp/agent-a \
OPENSQUILLA_GATEWAY_STATE_DIR=/tmp/agent-a/state \
  opensquilla agent -m "task A" --json --event-stream-stderr &
OPENSQUILLA_STATE_DIR=/tmp/agent-b \
OPENSQUILLA_GATEWAY_STATE_DIR=/tmp/agent-b/state \
  opensquilla agent -m "task B" --json --event-stream-stderr &
wait

On Windows, set both values in each child process environment. If an orchestrator copies config.toml or .env, remove or rewrite every state_dir or OPENSQUILLA_GATEWAY_STATE_DIR value as well. Explicitly shared session DB, workspace, scratch, transcript, and usage paths remain shared by design.

Development-only OpenTUI terminal UI. Release installs continue to use the Python-native chat. The richer full-screen frontend currently runs only from a Develop from source checkout; no companion host is published in release assets or installed by the release installer. From the checkout, install the pinned Bun dependencies once, then launch against that same source tree:

bun install --frozen-lockfile --cwd=src/opensquilla/cli/tui/opentui/package
OPENSQUILLA_TUI_DEV_SOURCE_HOST=1 uv run opensquilla chat --ui tui

Use opensquilla chat --ui plain to require the stable renderer. See docs/tui.md for terminal chat usage and docs/features/tui-frontend.md for backend details.

Open the Web UI at http://127.0.0.1:18791/control/. The Health view shows whether OpenSquilla is ready, what is not ready, and the next recovery steps. From the CLI, run:

opensquilla doctor
opensquilla doctor --json
opensquilla doctor --config ./opensquilla.toml --json

/health and /healthz are lightweight liveness endpoints for process checks. opensquilla doctor and the Web UI Health view are the readiness surfaces for provider config, memory, logs, search, channels, sandbox posture, router, image generation, and recovery guidance. Press Ctrl+C to stop a foreground gateway.

Other command groups include sessions, skills, memory, migrate, cron, channels, providers, models, and cost. Run opensquilla --help or opensquilla <group> --help for details.

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

opensquilla gateway run --listen 0.0.0.0 --port 18791

OPENSQUILLA_GATEWAY_IMAGE=ghcr.io/tokenrhythm/opensquilla:latest docker compose up -d

docker build -t opensquilla:local .

Provider tiers, sandbox tuning, image generation, and concurrency settings live in opensquilla.toml.example.

Per-version highlights live in CHANGELOG.md and docs/releases/.

Results from our technical report. Scores come from each benchmark's own grader, and costs use one provider price list across all frameworks. An OpenSquilla row is either a forced single-model run — the full harness with routing disabled, which isolates the harness contribution — or a multi-tier routing pool.

25 tasks covering file operations, data processing, web retrieval, creative output, tool use, and memory recall. Score is the cross-task mean; cost is the total for the run.

Forced single-model OpenSquilla takes the top score, +1.3 over OpenClaw's Opus-4.7 run at 22% lower cost. The Opus-4.7-backstopped routing pool holds 99.96% of the 92.55 baseline score at $0.69. The OpenRouter Auto row is a query-level routing baseline measured under a separate protocol and is not directly comparable to the other rows.

100 domain-specific business tasks across 13 domains, from clinical assistance and content operations to legal, HR, insurance, and real estate. Cost is average billed cost per task.

The harness alone adds 5.8 points to the same model, at higher cost — this is the one benchmark where the quality point sits above the baseline on price. The routing pool trades the gain back for a 37% cheaper task, landing near the GLM-5.1 direct run.

350 multilingual SWE-bench-style repair tasks from 8 languages and 43 repositories, with the task set, containers, prompts, turn limits, and timeouts all fixed. Cost is average billed cost per task.

Forced single-model GLM-5.2 resolves the most tasks: +5.1 points over OpenClaw's direct GLM-5.2 run, and ahead of OpenClaw's strongest single model at about 31% of its cost. The routing pool matches OpenClaw's GLM-5.2 run at half the cost, delegating 207 of the 350 tasks (59%) to the inexpensive DeepSeek-V4 Flash tier and escalating 139 to the top tier.

100 cross-domain deep-research tasks, graded on factual accuracy, completeness, objectivity, presentation quality, and citation quality. Cost is average billed cost per task; tokens are input plus output, in thousands.

The harness alone edges past OpenClaw's direct run while nearly halving cost; turning routing on keeps 99.94% of that score at 67% below the OpenClaw run. The routed run spends more tokens than the fixed one, so the saving comes from the price mix of the calls rather than from shorter prompts.

The high-accuracy mode drafts with several proposer models and fuses the drafts with an aggregator. Every row runs in the same OpenSquilla harness over DRACO, so the execution substrate is fixed and only the model allocation changes.

The ensemble outscores the strongest single model on both search providers: +1.02 points at 31% of its cost under DuckDuckGo, and +2.03 points at 90.8% lower cost under Brave. Fable 5 completed 94 of the 100 tasks under DuckDuckGo and 93 under Brave, and is scored on the tasks it completed; every other row completed all 100. Letting the router assemble the proposer set at run time, with only the aggregator fixed, comes within 0.51 points of the hand-picked configuration at 15.8% lower cost.

The quality is paid for in tokens and wall-clock time, which parallel proposer execution, early stopping, and per-proposer budgets control.

brew install libomp
opensquilla gateway restart

opensquilla onboard --provider openrouter --api-key-env OPENROUTER_API_KEY --router recommended
opensquilla gateway restart

OpenSquilla is inspired by OpenClaw. Bundled third-party content is attributed in THIRD_PARTY_NOTICES.md.

Community contributors are acknowledged in CONTRIBUTORS.md, including release-specific attribution notes for squash-merged or replayed work.

Thanks to all the people who contribute to OpenSquilla.

Contributions of every kind are welcome — bug reports, feature ideas, documentation, new provider or channel adapters, skills, and core runtime work. See CONTRIBUTING.md, then open an issue or pull request on GitHub.

Code of Conduct · Security · Privacy · Code signing policy · Third-party notices · Support · License (Apache-2.0)

If you use OpenSquilla in your research, please cite our technical report:

@misc{opensquilla2026,
  title         = {OpenSquilla: Token-Efficient Agent = Models + Routing Harness},
  author        = {{TokenRhythm Technologies}},
  year          = {2026},
  month         = aug,
  eprint        = {aixiv.260822.000001},
  archivePrefix = {aiXiv},
  howpublished  = {aiXiv preprint},
  url           = {https://aixiv.science/abs/aixiv.260822.000001},
  note          = {Version 1.0, under review}
}