Skip to content
Build Perch
Browse documentation

Where your data lives

Short answer: Your files, chats, task boards, and search index all live as ordinary files on your machine; there’s no Build Perch server in the data path and no telemetry. The AI models you choose to talk to are the exception: whatever you send to a remote provider (OpenRouter or Ollama Cloud) goes to that provider, under your own account.

What stays on my machine?

A workspace is a folder. Everything Build Perch manages inside it (conversations, task cards, generated documents, the local search index, undo snapshots) is written to a visible BuildPerch folder alongside your files, in ordinary formats you can open outside the app. Nothing is copied to a cloud account, and no Build Perch server can read your content back.

“No server in the data path” is specific: it means there is no telemetry and no analytics. The app doesn’t phone home to log what you typed, what files you opened, or how you used it.

What actually leaves your machine?

Two kinds of traffic leave the app, and they’re not the same thing:

  • The AI requests you make. When you send a message, the prompt (and any files or context you attached) goes to whichever provider that chat is connected to (OpenRouter or Ollama Cloud, both remote services under your own credentials), or stays on your machine entirely if the chat is running a local Ollama model. Build Perch doesn’t sit between you and that provider’s own privacy policy; it’s a pass-through you control per chat.
  • A small set of app-infrastructure calls that never carry your content. The app itself contacts a small, fixed set of endpoints: our licensing service (activation, license checks, and trial issuance), the signed-update feed, and (only if you turn on local transcription) a one-time speech-model download from a public model host. The licensing calls carry only a license key and a one-way-hashed machine fingerprint (plus the OS name and a generic device label at activation, and the app version when starting a trial); the update check and the model download carry no identifying data at all. None of them ever carry your files, chats, prompts, or your computer’s actual name. See Trial, licensing & devices for the exact fields.

If the app crashes, the details are written to a local crash log automatically and shown to you the next time you launch. That safety net is always on, and the file stays on your machine unless you choose to share it.

How an AI request is put together

When you send a message, Build Perch assembles one request on your machine in three layers, then sends it to the chat’s provider:

  1. A system prompt: the standing setup the app builds automatically each turn: who the AI is, the rules it follows (including that anything it reads from a tool, file, or web page is data to analyze, not instructions to obey), the behavior of the mode you’re in, and descriptions of the tools and skills you’ve turned on. Your workspace memory, a file-tree summary, or your task cards are included too, but only when the matching skill is on.
  2. The conversation so far: your earlier messages and the AI’s replies. Once a chat runs long, older turns are replaced by a running summary while recent ones stay word-for-word; nothing is deleted from the chat itself.
  3. Your new message: what you typed, plus any files, images, or anchored prompt presets on the chat.

You don’t normally see the system prompt itself, but the context meter’s popover shows how large it is.

What a request costs

Providers charge by the token, and every turn re-sends the system prompt, plus the conversation (or its summary), plus your message. So a turn costs more when more skills are on (more tool descriptions), when a large memory file or file tree is injected, or when the conversation is long, which is what the summarizer keeps in check. On providers that support prompt caching, the unchanging system-prompt layer is served from the provider’s cache instead of being reprocessed, which keeps a long conversation cheaper; the context meter’s “Prompt cache” line shows how much of your last send was cached. There’s no aggregate spend dashboard; cost shows as each model’s per-million-token rate, and you pay your provider directly. See The context meter.

Good to know

  • Local features stay local regardless of which provider a chat uses: semantic search (embeddings run on-device with a bundled model), voice dictation and transcription (local speech models), and media editing (bundled media tools) never send your files anywhere.
  • Media generation (images, speech, music, video) is the one feature that always requires a remote call; it’s OpenRouter-only. Editing existing media stays local.
  • Privacy Mode further restricts what OpenRouter itself does with the requests you do send it.