Skip to content

Zyquo Agent 1.0.0

Latest
Latest

Choose a tag to compare

View all tags
@spboucher-ai spboucher-ai released this 30 Jul 22:35
v1.0.0
6829b5c

Zyquo Agent 1.0.0 — first public release

Zyquo Agent is a native macOS app that turns cloud LLMs into an agent that actually operates your Mac — it runs real bash commands, drives apps with AppleScript, reads and writes files in isolated per-task workspaces, and keeps iterating until the job is genuinely done. Every action passes a safety gate that stops and asks you before anything risky.

⬇️ Install

  1. Download ZyquoAgent.dmg below.
  2. Open it and drag Zyquo Agent to your Applications folder.
  3. Launch it — no Gatekeeper warning and no right-click workaround: the app and the DMG are both signed with a Developer ID, notarized by Apple, and stapled.
  4. Open Settings → Providers & Keys (⌘,) and paste at least one API key. Zyquo Agent ships with none and talks only to providers you configure.
$ spctl -a -vv "/Applications/Zyquo Agent.app"
/Applications/Zyquo Agent.app: accepted
source=Notarized Developer ID

Requirements: macOS 13.0 Ventura or later · Universal binary (Apple Silicon + Intel) · one BYOK API key.

✨ What's in it

A real agentic loop — plan → act → observe → reflect, driven by native tool calling and looping until the model answers without needing tools. It drafts a live checklist you can edit mid-run, re-plans after failures, and verifies its own work with tools before declaring done.

Tools that do real work — bash (line-streamed output, timeouts, exit codes, cancellable), osascript for macOS automation, and workspace-scoped read_file / write_file / edit_file / list_dir / search_files.

Safety as a headline feature, not an afterthought

  • Three modes per task: Manual, Guarded (default), Autonomous.
  • Commands are parsed, not pattern-matched: compound lines split on &&, ||, ;, |; $(…) substitutions classified too; the verdict is the most severe subcommand.
  • Always-ask circuit breakers in every mode, Autonomous included: any sudo, rm -rf outside the workspace, curl | sh, launchctl, writes to system paths, and more. sudo never runs silently.
  • Inline approval cards showing the exact command, risk, cwd and intent — Approve / Approve & Remember / Edit / Deny.
  • Append-only audit log of every executed action, exportable.

Long runs don't fall over — token accounting with compaction at 85% of context (plan, MEMORY.md, and the last 6 steps stay verbatim), tool outputs over 8 KB offloaded to files with in-context references, and a LoopGuard that enforces step/token/time budgets and pauses to ask rather than spinning.

A command center, not a chat box — task sidebar with live status pills, step-card transcript with streaming output, plan panel with budget meters, and an activity drawer with Live / Audit Log / Files tabs.

170 models across 12 providers — OpenAI, Anthropic, xAI, Mistral, Gemini, Qwen, DeepSeek, Kimi, Perplexity, Together, DeepInfra, Cerebras, plus custom OpenAI-compatible endpoints. Keys live in an AES-256-GCM encrypted vault (deliberately not the Keychain).

Productivity — 29 task templates with variables, personas, ⌥Space Quick Task panel, ⌘K command palette, Markdown/PDF transcript export, full-text search, and a menu bar extra.

🔬 Verified, not asserted

Suite Result
Provider tool-calling, live, every agent-capable model 77 / 80 green
End-to-end scenarios in real workspaces 11 / 11 pass
Safety tests, live 8 / 8 pass
PolicyEngine self-check 38 / 38

Full reports: VERIFICATION.md · EVALUATION.md. The 3 non-green models fail for documented external reasons (DeepInfra strips Gemini's required thought_signature; kimi-k3 is capacity-throttled upstream).

📌 Known limitations

  • Two safety settings are persisted and shown but not yet enforced by dedicated engine hooks: requireApprovalForAppleScript in Autonomous mode, and workspaceEscapePolicy = deny. The engine already asks for AppleScript in Manual/Guarded and always asks on workspace escape, so the defaults are safe.
  • Auto-generated task titles can be clipped for very short prompts.
  • swift test needs an XCTest-capable toolchain; --verify-policy, --run-mock and --run-ui-smoke are the equivalents that run anywhere.

🙏 Notes

Built entirely with Swift Package Manager and the command-line toolchain — no Xcode IDE. Bring your own API keys; nothing is bundled and no telemetry is collected.

Author: Simon-Pierre Boucher · Contact: contact@spboucher.ai

Assets 3
Loading