Parad0x Labs has released VOOL 0.7.7-beta, a local-first personal agent that distinguishes itself by making every action offline-verifiable through signed receipts. Unlike competitors that rely on cloud-based trust models, VOOL records each tool call with what was asked, what ran, and what returned, allowing users to verify integrity without an internet connection. The project is MIT-licensed, requires no account, and supports both local models via Ollama and bring-your-own-key cloud setups.

Memory Efficiency and Benchmark Performance

In head-to-head testing using the LoCoMo dataset, VOOL demonstrated significant advantages over established agents like OpenClaw and Hermes. With the same GPT-4o grader and model configuration, VOOL achieved an 82.7% right answer rate compared to OpenClaw’s 71.3% and Hermes’ 62.7%. More critically, VOOL consumed only 9.6k tokens per question, a 5.4x reduction compared to OpenClaw’s 51.8k. This efficiency translated to a cost of $0.0033 per question, roughly one-third of OpenClaw’s $0.0100 cost, while also delivering a faster median answer time of 11.2 seconds.

Cryptographic Proof and Local Control

The core differentiator for VOOL is its 'honesty receipt' system. Users can run python -m core.honesty_receipt verify-last to inspect the most recent session ledger, verifying that the signed record's integrity remains intact. This approach shifts the paradigm from trusting the agent’s output to verifying its process. The agent stores facts and decisions across sessions in readable files, giving users direct control over memory retention and deletion. While the receipts verify integrity, the documentation explicitly notes they do not guarantee the correctness of every model answer, a crucial distinction for security-conscious developers.

Platform Availability and Installation

VOOL 0.7.7-beta currently supports macOS (Apple Silicon and Intel) and Linux x86_64, with Windows available only via experimental source installation. The macOS beta is not yet notarized by Apple, requiring users to manually approve the app in Privacy & Security settings on first launch. Installation from source requires Python 3.11 or newer and includes a bootstrap script that handles hardware checks and local model setup. Users can choose between 'local-only' or 'local-max' profiles to balance resource usage, though local inference on Intel Macs may be too slow for certification tasks.

Key Takeaways

  • VOOL outperforms OpenClaw and Hermes in memory benchmarks with 82.7% accuracy and 5.4x lower token usage.
  • Signed receipts allow offline verification of agent actions, enhancing transparency and trust.
  • The agent is MIT-licensed, free, and supports both local Ollama models and cloud providers without requiring an account.
  • Current beta limitations include lack of Apple notarization and slow local inference on Intel Macs.

The Bottom Line

VOOL’s combination of statistical superiority and cryptographic transparency sets a new standard for local agents, though its current beta status and hardware limitations mean it is best suited for developers willing to tinker before it is ready for mainstream adoption.