Help & support.

No. Ingestion, hybrid search, cited chat, OCR, transforms, the MCP server, and the optional workspace timeline all run on your Mac on the Apple Silicon GPU. Nothing you index ever leaves your machine — there's no account, no telemetry, and no analytics SDK. Your vaults live inside the app's on-device container. The only requests Silo Vault can make are ones you turn on: the consent-gated model download and the optional model-catalog check below, neither of which carries any of your content.

It is off by default and never runs until you turn it on. Once enabled, Silo Vault periodically downloads a small signed model-catalog file from ohmslaw.net — a site operated by Silo Systems and hosted on Vercel — so it can tell you when a newer local model build exists. It is a one-way download of metadata: no document, vault name, search query, chat message, or account identifier is sent, and there is no account to identify you with. The request exposes only what any HTTPS request exposes — your IP address and a user-agent string. Silo Vault logs nothing about it, though the hosting infrastructure may keep standard web access logs. Turning the setting off stops the checks, and if the catalog is missing, expired, or fails signature verification, Silo Vault just keeps using the model list it already knows.

It means the model has not been tested on Silo Systems hardware. Some models the catalog offers are marked this way so that large builds can reach machines we do not own. Their memory and disk requirements are estimated, not measured — calculated from the model's published size and format rather than from a real run — and Silo Vault shows you the basis for the estimate before anything downloads. An unvalidated model is never picked for you: it is never a default at any memory tier, and installing one takes an explicit extra confirmation on top of the normal model-download consent. It may need more memory than estimated, or run poorly, on your Mac. If a real evaluation is done later, the same model is republished as a tested one with measured requirements.

Each vault can run as a local Model Context Protocol server that speaks stdio only — no open port. In Silo Vault, enable MCP for the vault you want to expose; it gives you the command to launch the local server. Add that command to your tool's MCP configuration — for Claude Code, register it as an MCP server (claude mcp add, or an entry in your MCP config); for Cursor, add it under Settings → MCP; for Codex, add it to your MCP server list. Once connected, the tool queries your indexed documents on-device and gets cited answers back. No network port is opened and no remote connection is made.

Workspace recall is optional and off by default — it never runs until you explicitly turn it on. When enabled, it builds a private, searchable timeline from periodic on-device screen OCR and optional audio transcription. Everything is processed and stored on-device only; nothing is transmitted. A menu-bar recording indicator stays visible the entire time capture is active, you can exclude specific apps, secure fields (like password entry) are excluded automatically, and you set a retention window after which old entries expire. You can clear the entire history or disable the feature at any time.

PDF, DOCX, PPTX, XLSX, HTML, CSV, Markdown, and plain text parse locally. Scanned PDFs and images are read with on-device Apple Vision OCR via the bundled silovault-ocr CLI — no cloud OCR. Everything is indexed into the vault you choose.

Silo Vault uses hybrid RAG: SQLite FTS5 keyword search fused with local EmbeddingGemma 300M vectors via Reciprocal Rank Fusion, so exact terms and fuzzy meaning both surface the right passage. Your selected Gemma 4 E2B, Gemma 4 E4B, Gemma 4 12B, Gemma 4 26B MoE, or Gemma 4 31B model then writes the answer locally, and every answer cites its sources with page numbers so you can verify it against the original file.

Only for model setup, links you choose, and the optional model-catalog check if you enable it. After explicit consent, Silo Vault downloads revision-pinned Gemma 4 E2B, Gemma 4 E4B, Gemma 4 12B, Gemma 4 26B MoE, or Gemma 4 31B chat weights and EmbeddingGemma 300M weights from the disclosed Hugging Face hosts, verifies them, and then runs ingestion, search, chat, OCR, MCP, and workspace recall offline.

It's a per-vault preset for reviewing agreements on-device: risk flags, clause extraction, a risk-scoring dashboard, negotiation-playbook comparison, clause-level compare/redline, and one-click PDF export. It is informational document review only — not legal advice, and no attorney-client relationship is created.

They are coming in version 1.8.0 and are not in the current App Store download. The update exports selected vaults as authenticated AES-256-GCM .silovault archives, verifies integrity and previews naming conflicts before import, and adds exact citation spans, coverage scoring, unsupported-claim flags, and abstention when evidence is weak.