Every file Claude Code reads in a session gets sent to Anthropic's API as input tokens. That includes personal notes, health records, financial data, and private journal entries. There is no local processing mode. No technical workaround changes this.
That said, Anthropic's paid API plans don't use your inputs for model training by default. Your notes pass through their servers but aren't retained for training. See Anthropic's privacy policy for the full terms.
Kepano (Steph Ango, CEO of Obsidian) put it bluntly: he has no desire to upload his personal unencrypted vault data to Anthropic or anyone else. If the person who built the tool draws that line, you should at least think about where yours is.
My recommendation: keep a dedicated work vault that you point Claude Code at, and a separate personal vault that stays offline. If you use a single vault, scope access to a work/ subfolder using the filesystem MCP with a restricted path:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/vault/work"]
}
}
}
Replace the path with your work subfolder. Claude Code can only read and write inside that path. Your personal folders stay completely out of reach.