Can an AI Assistant Search Your Documents Without Uploading Them? What Stays on Your PC and What Doesn't
"Without uploading" is true, and it is also where most explanations stop too early. When an AI assistant searches your documents through a local server, no file is sent anywhere as a file. But the assistant reads what it finds, and what it reads becomes part of the conversation. This post draws that line precisely, using LocalSynapse as the example because we can state exactly what it does.
What stays on your PC?
- Your documents. LocalSynapse reads them where they are.
- The index. Extracted text and search data are stored in a local database on your machine.
- The AI model, if you install it. Semantic search uses the BGE-M3 model running on your CPU. It is optional (about 2.3 GB); keyword search works without it.
- The MCP server. When Claude Desktop or another client uses LocalSynapse, it starts LocalSynapse as a local process on your PC. The search runs there.
What leaves your PC, and when?
| What | Where it goes | When |
|---|---|---|
| Tool results the assistant reads (passages, extracted text, spreadsheet cells) | Your assistant's provider, if it is a cloud assistant | Each time the assistant calls a tool |
| Anonymous usage report: random install ID, app version, OS, language, usage counts, technical health metrics | LocalSynapse | Each time the app starts, unless turned off |
| Update check | GitHub | At startup, unless turned off |
| AI model download | From Hugging Face | Only if you choose to install the model |
| Diagnostics report (no install ID) | LocalSynapse | Only when you preview it and press send |
File names, folder paths, file contents, email contents and your search text are not part of LocalSynapse's own reports. To stop the usage report and the update check, open Security in the app and turn off External communication. Details are on the privacy page.
The first row is the one that matters most, and LocalSynapse does not control it. Whether tool results leave your PC is decided by the assistant you connect. A cloud assistant sends them to its provider just as it would send text you pasted.
How do I keep what the assistant reads small?
- Index only what you'd share. LocalSynapse indexes the folders you choose. A folder that isn't indexed can't be found through MCP.
- Prefer passages over whole files.
search_filesreturns the matching passages and their position, not the whole document. Itspurposesetting controls depth:locatereturns short passages from up to 20 documents,answerlonger ones from 8,surveythe shortest from 50.get_file_content, by contrast, returns a file's full extracted text, so ask the assistant to quote from search results when that is enough. - Narrow spreadsheet reads.
get_spreadsheettakes a sheet index and a cell range such asA1:Z100, so the assistant can read one table instead of a whole workbook (how it works).
Can the assistant change or delete anything?
No. The nine tools search, list and read. The one removal-related tool, preview_index_removal, is read-only: it shows which index entries a cleanup would select, and the removal itself can only be confirmed in the desktop app. No tool edits, moves or deletes your files.
How does this compare with uploading a file to a chat?
Uploading sends the whole file to the service in one go, and you choose each file by hand. With a local index, the assistant searches across everything you indexed and fetches only the passages it needs. That means it can find documents you forgot you had — and it means the passages it fetches are the ones that leave. Neither approach is "private" in the sense of nothing leaving, unless the assistant itself runs locally.
What LocalSynapse does not do
- It doesn't make a cloud assistant private. It controls what is indexed and how much a tool returns, not where your conversation goes.
- It skips cloud-only placeholder files (for example OneDrive online-only files) instead of downloading them.
- No OCR, so scanned PDFs without a text layer are not searchable.
- Windows and macOS only, on the CPU. The Windows build is not code-signed and the macOS builds are not signed by Apple. The source code is not published.