Restore gdrive server with search file ID fix and download tool

[aaronpolhamus]

Summary

• Restores the archived src/gdrive/ server (removed in May 2025)
• Fixes search results to include file IDs — the Drive API fetches files(id, name, mimeType, ...) but the output only included name and MIME type, making results unusable for downstream operations
• Adds a download tool that saves files to a local temp directory instead of returning base64 blobs via the resource handler, which overflow LLM context windows
• Fixes OAuth token auto-refresh — the server was creating new google.auth.OAuth2() without client credentials, so the refresh_token was unusable once the access_token expired (~1 hour). Now reads client_id/secret from the OAuth keys file.
• Makes MCP resources opt-in via GDRIVE_ENABLE_RESOURCES env var — some MCP clients (e.g., Gemini CLI) call resources/list during initialization, which triggers drive.files.list() and can hang. Default is tools-only mode (search + download) for cross-client compatibility.

Motivation

The gdrive server is essential for LLM workflows that need to search and read Drive files (especially PDFs, images, and other binary formats). Two prior community PRs attempted similar improvements:

• Add gdrive_read_file tool and improve search results #1092 — "Add gdrive_read_file tool and improve search results"
• Refactor Google Drive Server: Transition to Tool-Based Architecture with Enhanced File Handling #1353 — "Transition to Tool-Based Architecture with Enhanced File Handling"

Both were closed when the server was archived. The underlying issues remain: without file IDs in search output, clients can't act on results; without a download tool, binary files either overflow context (base64 blobs) or are inaccessible.

Changes

1. Search output includes file IDs

// Before:
`${file.name} (${file.mimeType})`

// After:
`${file.name} (${file.mimeType}) [id:${file.id}]`

2. New download tool

• Takes a fileId (from search results) and saves the file to a local directory
• Google Workspace files are auto-exported: Docs → Markdown, Sheets → CSV, Presentations → Plain text, Drawings → PNG
• Binary files (PDFs, images, etc.) saved as-is
• Download directory configurable via GDRIVE_DOWNLOAD_DIR env var (default: /tmp/gdrive-downloads)
• Filenames are sanitized for filesystem safety

3. OAuth token auto-refresh fix

The original server created new google.auth.OAuth2() without client_id or client_secret. This meant the refresh_token in the stored credentials was unusable — when the access token expired (~1 hour), Google's token endpoint returned invalid_request because the refresh grant requires client credentials. Now reads client_id/client_secret from the same OAuth keys file used during the initial auth flow (GDRIVE_OAUTH_PATH) and passes them to the OAuth2 constructor.

4. MCP resources opt-in (GDRIVE_ENABLE_RESOURCES)

Some MCP clients invoke resources/list during initialization, which triggers drive.files.list() and can hang or timeout. Resources are now disabled by default — the server advertises only tools capability. Set GDRIVE_ENABLE_RESOURCES=true to re-enable ListResources and ReadResource handlers.

This is a clean capability negotiation: the server honestly advertises what it supports based on the env var, and the MCP SDK validates that handlers match declared capabilities.

5. Standalone build

• tsconfig.json is self-contained (no longer extends root tsconfig)
• Version bumped from 0.6.2 → 0.7.0

Test plan

• TypeScript compiles cleanly with npm run build
• search returns file IDs in output format [id:...]
• download saves binary files (PDFs) to local temp directory
• download exports Google Docs as markdown
• Filename sanitization handles special characters
• OAuth auto-refresh works after access token expiry (tested with expired token + valid refresh_token)
• Server starts in tools-only mode (default, no GDRIVE_ENABLE_RESOURCES)
• Server starts with resources enabled (GDRIVE_ENABLE_RESOURCES=true)
• Tested with Claude Code and Gemini CLI — both connect and use search/download tools successfully

Notes

We've been running these fixes in production for Claude Code and Gemini CLI MCP workflows (searching and reading ~300 signed compliance PDFs on Google Drive). Happy to maintain this server going forward.

Contact: aaron@mivest.io

🤖 Generated with Claude Code