Skip to content
Build Perch
Browse guides

How to use AI on your files without uploading them

Short answer: Point Build Perch at a folder on your disk (that becomes a workspace) and the AI reads, searches, and edits those files directly from disk, with no separate cloud store to upload them into first. Connect a local Ollama model and the content never leaves your machine at all; with a cloud model, only what you actually send in a message reaches that provider, under your own account.

Steps

  1. Open a workspace over the folder you want the AI working in. To bring in an existing folder, use Import (in the workspace toolbar or the Manage Workspaces window). To start fresh, use New Workspace: it creates a new folder inside your root’s Workspaces folder. Either way, Build Perch reads and writes only inside that folder; nothing is copied elsewhere to be indexed or read.
  2. Ask about a file in chat. The AI uses tools like read_file, find_text, and search_files to read or search your files on disk in that moment; each call shows up as a row in the Tool Activity log, so you can see exactly what it opened.
  3. For a hard guarantee that file content never crosses the network, connect a local Ollama model instead of a cloud one (see How to run AI models offline on Windows 11). Nothing about the conversation, including any file content in it, is sent anywhere.
  4. If you’re using a cloud model (OpenRouter) instead, leave Privacy Mode on (Integrations → OpenRouter; it’s on by default) so anything you do send only routes to providers that attest they don’t store or train on it.
  5. Right-click any folder you want off-limits entirely and choose Lock from AI: enforced by the app on every read, write, and even local search hit, regardless of which provider the chat is using.

Notes & gotchas

  • Opening a workspace doesn’t dump its contents into a chat. The AI only sees a file’s content when a tool call actually reads it, or when you attach or paste it yourself.
  • Local semantic search runs entirely on-device: file embeddings are computed with a bundled local model, not a hosted one, so indexing your files for search never involves a network call, no matter which chat provider you’re using.
  • “Without uploading” has one honest limit: if a chat is connected to a cloud provider (OpenRouter or Ollama Cloud) and the AI reads a file to answer you, that content is sent to that provider for the turn, the same as pasting it in yourself. Only a fully local Ollama chat avoids that entirely.
  • Privacy Mode is an attested restriction on OpenRouter’s routing, not a guarantee Build Perch can verify end-to-end; see Privacy Mode (ZDR) for the exact framing.