GitHub - Parad0x-Labs/vool: VOOL — the local-first personal agent runtime by Parad0x Labs: local LLM by default, memory, tools, permissions and signed receipts. macOS beta.

The open-source agent that remembers more, spends less, and runs on your machine.

VOOL is a local-first personal agent from Parad0x Labs. It reads your projects, edits files, runs tests, researches the web, and remembers what matters for next time. Free, MIT-licensed, no account. Start on a local model with no cloud key, or bring your own provider.

Beats OpenClaw and Hermes on memory

Same 150 never-used LoCoMo questions. Same model answering for all three agents. Same GPT-4o grader.

• +11 points over OpenClaw, +20 over Hermes, and the fewest confidently wrong answers (18 vs 31 and 35). Not luck: p = 0.005 and p < 0.001 on matched questions.
• 5.4x fewer tokens than OpenClaw, 3.5x fewer than Hermes. A third of OpenClaw's cost per answer.
• In the app today. The memory shipped in VOOL 0.7 scored 123 / 150 on the same questions, level with the build above.
• Every run is public, including the round where VOOL's lead was too small to count. See the proof

Why VOOL

• It does the work. Reads, plans, calls tools, checks results, and keeps going: code, files, commands, documents and the web. Tools and permissions
• It remembers. Facts and decisions carry across sessions, stored as readable files you can inspect and forget on demand. How memory works
• Your models. Local models through Ollama and supported runtimes, offline after setup, or a cloud model with your own key. Local and cloud
• Proof, not promises. Every tool call is recorded with what was asked, what ran and what came back; signed receipts verify offline. Receipts
• You stay in charge. Workspace scopes, approval modes and spending limits decide what the agent may do. Spending limits

Get started · Documentation · Benchmarks · Explore the code · Community

Install

Current version: 0.7.7-beta. Download the app for your computer from the v0.7.7-beta release, or install from source with the bootstrap script below.

Downloads

Each download has a SHA-256 checksum on the release page. Check it before you open the file (how). Not sure which Mac you have? Open the Apple menu and choose About This Mac: "Apple M…" means Apple Silicon, "Intel" means Intel. The macOS app needs macOS 14 Sonoma or newer.

First open on a Mac. This beta is not notarized by Apple yet, so macOS asks you to confirm the first time:

1. Open the DMG and drag VOOL into Applications.
2. Double-click VOOL. macOS says it can't check the app; click Done (or Cancel).
3. Open System Settings → Privacy & Security, scroll down, and click Open Anyway next to VOOL. Confirm with your password. (On macOS 14 you can instead Control-click VOOL and choose Open.) You only do this once.

Linux. Unpack the tarball and run its installer; no sudo is needed. Full steps are in the install guide.

Using a local model? Before a local model can write answers, open Settings → Models & Providers → Local models and press Certify for it once. This runs a short check on your machine.

From 0.7.5 on, VOOL updates itself: with "Go online on its own" on, it shows a notice and installs a new version with one press. Updating from 0.7.4 or 0.7.3 on a Mac: turn on "Go online on its own" (Settings > Privacy & Permissions) and press Check for updates once. From 0.7.4 or older on Linux, or 0.7.2 and older on a Mac: download 0.7.7 and install it over your current version. Your chats and settings live in your profile, outside the app. Release notes: CHANGELOG. Release state, signing and updates: release status.

For model choices, system requirements, checksums, and troubleshooting, see the install guide and Windows capability matrix.

From source: macOS / Linux

Download the bootstrap script, then run it:

curl -fsSLo bootstrap_vool.sh https://raw.githubusercontent.com/Parad0x-Labs/vool/v0.7.7-beta/installer/bootstrap_vool.sh
bash bootstrap_vool.sh --ref v0.7.7-beta

From source: Windows PowerShell

Invoke-WebRequest https://raw.githubusercontent.com/Parad0x-Labs/vool/main/installer/bootstrap_vool.ps1 -OutFile bootstrap_vool.ps1
powershell -ExecutionPolicy Bypass -File .\bootstrap_vool.ps1

The installer prepares the Python environment, checks your hardware, sets up the local model, and launches the local services. Installing from source needs Python 3.11 or newer. If none is found, the installer tries to set one up. The app downloads include their own Python. Allow time and disk space for the initial model download. Local model inference works offline after setup; web tools and cloud models need a connection. Legacy integration cleanup is tracked in the repository scope map.

bash bootstrap_vool.sh --ref v0.7.7-beta --install-profile local-only
# Or use the larger local model profile:
bash bootstrap_vool.sh --ref v0.7.7-beta --install-profile local-max

python -m apps.vool_cli install-profile --set local-max

py -m apps.vool_api_server

Give it something real to do

Once your model and workspace are configured, try:

• Code: “Find why this test fails, fix the cause, and run the affected tests.”
• Research: “Compare these options using current sources and save the findings.”
• Documents: “Read these PDFs and summarize the differences with references.”
• Memory: “Remember this project decision so we can pick it up next session.”

These are starting prompts, not fixed demos. Available tools depend on your setup and permissions; task quality depends on the model you choose. Web research requires web access to be enabled.

Inspect the local service at /healthz, its activity at /trace, and configured features at /api/runtime/capabilities.

Check the receipts

From your installed environment, run the offline receipt demonstration:

python -m core.honesty_receipt demo

It creates and verifies sample receipts, then demonstrates tamper detection. To inspect the most recent real session ledger:

python -m core.honesty_receipt verify-last

A signature verifies the signed record's integrity; it is not a guarantee that every model answer is correct.

Memory benchmarks

The full write-up has every round, the setup, the spend and the weak spots: VOOL is slowest to ingest a conversation (it embeds every turn on your machine), and multi-hop questions that join facts from several sessions are hard for every agent tested. Memory benchmarks.

Explore further

Meet the related ecosystem projects, including Dark Null Protocol for privacy-preserving settlement.

VOOL also includes optional wallet/payment integrations and bounded local helper orchestration. Peer mesh, DHT and Brain Hive are retained experimental research, outside normal production networking. See the repository scope map for current runtime, integration, research and compatibility boundaries. Wallet and spending features require their own setup and authorization.

Built by Parad0x Labs. Contributor: sls_0x.