v0.6.4

Bug Fix

• Fix update_record content writes for non-plaintext records — update_record previously hard-coded writes to the plain text property, which silently failed (or wrote to the wrong field) for markdown, html, webarchive, rtf, and rtfd records. The bridge now detects the record type at AppleScript runtime and writes to the correct property:
  • markdown → markdown source
  • html / webarchive → source
  • rtf / rtfd → rich text
  • everything else → plain text

v0.6.3

Bug Fix

• Fix crash on records with no tags — get_record (and other calls using parseRecord) would crash the server process silently when querying a record with zero tags, due to an invalid 1...0 closed range in Swift. This caused "Connection closed" (-32000) errors on the MCP client side.
• Fix same crash pattern in parseCustomMetadata() for records with no custom metadata.
• Add descriptor validation — parseRecord() now validates the AppleScript descriptor has the expected number of fields before parsing, returning a safe fallback instead of crashing on malformed data.
• Add stderr logging — AppleScript errors and malformed descriptors are now logged to stderr for diagnostics.

DOI downloads from sci-hub

When working with dt-mcp and interactively searching for scientific documents, you can now use the DOI to find your doc in DEVONthink or, when missing, download from sci-hub automatically.

• When searching for a scientific document by DOI, the AI first checks DEVONthink for existing copies
• If not found, uses sci-hub to download the PDF (tries multiple mirrors: sci-hub.se, .ru, .st, .ee, .ren)
• Parses the sci-hub page to extract the PDF URL (supports embed, iframe, onclick, object tags)
• Downloads the PDF and imports it into DEVONthink with metadata

Privacy Mode, Database Exclusion & Image Preview

• Privacy Mode - Global toggle to strip metadata from all documents
• Database Exclusion - Hide entire databases from AI access completely
• Image Preview - Analyze images with confirmation flow
  • Two-step flow: metadata/EXIF first, then image data
  • EXIF stripping (GPS, camera, author removed)
  • Adaptive compression to fit Claude Code limits
  • Private image modes: thumbnail, text_only, blocked

PRIVATE tag PII anonymization

Universal build.

Set tag 'PRIVATE' on any document to ensure the MCP server anonymizes document contents before it goes to LLM server. See readme for more details.