Home

ArkDrive Wiki

ArkDrive is a model-agnostic GUI computer-use executor for Linux/X11: a hardened runtime that sits between any tool-calling LLM and a real desktop, and makes GUI work accountable — per-action deadlines, fallback input ladders, structured dispatch receipts, progress sensing, and honest structured aborts instead of silent failures.

The driving model is bring-your-own (any OpenAI-compatible, tool-calling endpoint). Benchmarks are published per model, so the executor's quality and the model's quality are never confused.

Pages

• Test-Results — the full benchmark suite: all 10 task cards, per-model results, core/smoke test status, and the fun runs (Doom, Blender).
• Architecture — how the executor works: the loop, deadlines, input ladders and receipts, micro-agents, progress sensing, completion discipline.
• Safety-Model — kill switch, dry-run, audit log, human-in-the-loop points, deployment guidance.
• Hardening-Story — how the executor was hardened: 125 adversarial review passes, ~250 defects, one structural rebuild, and the rules that came out of it.
• Compatibility — who can drive ArkDrive: executor model presets (llama / OpenAI / DeepSeek / OpenRouter / Anthropic) and external coding agents (Codex proven; Claude Code and generic agents by contract).

In the repo

• README.md — pitch, quickstart, short benchmark table
• docs/ — the full architecture, safety, runbook-pack, and hardening documents (this wiki summarizes them)
• bench/ — task cards, runner, and machine-readable results