A local-first desktop agent for object storage on S3-compatible systems — operations, analytics, and management. It diagnoses issues, analyzes access logs and inventory, reviews account and bucket configuration, triages errors, profiles usage and cost, and recommends optimizations. Everything runs on your machine; your data and credentials never leave your computer.
It is evidence-driven and human-in-the-loop: the agent investigates with read-only tools and grounds its conclusions in artifacts you collected. It never mutates your storage. Cloud-side data-moving steps (downloads, large scans, evidence import from a bucket) always wait for your confirmation; a file you attach in the chat is local, so the agent analyzes it inline. Read-only investigation always runs autonomously (see Safety model).
Download the installer for your platform from
GitHub Releases.
Each release also includes SHA256SUMS-*.txt so you can verify your download.
| Platform | Asset |
|---|---|
| macOS (Apple Silicon) | …-macos-arm64.dmg |
| Linux (x64) | …-linux-x64.deb |
| Windows (x64) | …-windows-x64-setup.exe |
These builds are ad-hoc signed, not notarized / not Authenticode-signed, so each OS warns on first launch. This is expected — here's how to open it.
Because the app isn't notarized, macOS quarantines it after download and may say either "Apple cannot check it for malicious software" or "is damaged and can't be opened." Neither means the app is broken — it's the missing Apple signature. Pick one:
-
Terminal (most reliable). After moving the app to
/Applications, clear the quarantine flag and open it:xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app" open "/Applications/Storage Agent Workbench.app"
-
Finder. Right-click the app → Open → Open in the dialog. (If macOS says "damaged", use the Terminal step above — right-click Open won't clear that.)
The app starts in a few seconds. You'll see Sidecar: Connecting briefly while the local backend comes up.
Secrets are kept in an encrypted local vault, so there is no system keychain authorization prompt on any platform. After updating to a build that introduced the vault, re-enter your model API key and cloud credentials once (they aren't migrated from the old keychain) — you won't be prompted again.
sudo apt install ./storage-agent-workbench-*-linux-x64.debA WebKitGTK runtime is required (pulled in as a dependency on most distros).
Run …-windows-x64-setup.exe. SmartScreen may warn because the installer isn't
signed — choose More info → Run anyway. WebView2 is required (preinstalled on
current Windows; the installer fetches it if missing).
- Configure a model provider and an S3-compatible cloud provider.
- Discover the account and its buckets.
- Review bucket configuration — security, lifecycle, observability, cost.
- Import evidence (inventory / access logs) with explicit confirmation.
- Analyze inventory and access logs locally with DuckDB.
- Triage S3 / object-storage errors.
- Keep the investigation in a session (rename / pin / archive / delete / fork / search) and generate Markdown reports.
The interface is a thread-first agentic workbench: a session rail, a conversation thread where runs and findings render as inline cards, and a settings drawer. Dark and light themes; English and 中文.
- Local-first. App data lives in the OS app-data directory; nothing is sent anywhere except the cloud/model providers you configure.
- Secrets in an encrypted local vault. Access keys, secret keys, session tokens, and model API keys live only in an AES-256-GCM vault on your device (key protected per-OS) — never in SQLite, logs, reports, or model prompts.
- Read-only. No destructive or mutating S3 operations; no generic shell or arbitrary subprocess tool. The agent is a fully autonomous read-only investigator — it runs its read-only tools itself (no autonomy toggle).
- You confirm anything that moves cloud data. Downloads, large/full scans, and evidence import from a bucket always wait for your confirmation; there is no write tool. A file you attach in the chat is local, so the agent analyzes it inline without a separate confirmation.
- Agent context is bounded and sanitized; chain-of-thought is never persisted.
See docs/security.md for the full model.
- docs/install.md — installing per platform.
- docs/product.md — product shape and core jobs.
- docs/architecture.md — how the pieces fit together.
- docs/security.md — secret handling and safety rules.
- docs/signing.md — macOS signing/notarization status.
- docs/release.md — release flow and platform support.
- CHANGELOG.md — release notes.
Apache License 2.0. You may use, modify, and distribute this software,
including commercially, provided you preserve the copyright and NOTICE
attribution; the license also includes an explicit patent grant.