Google Drive MCP lets the agent reach into your Drive without you having to download files first. It searches by name, opens Docs / Sheets / Slides as text, and pulls PDFs as content the model can reason about. Read-only by design — no accidental overwrites of the contract you’ve negotiated for three weeks.
What it produces: search_files, read_file. Files returned as plain text where convertible (Docs → Markdown, Sheets → CSV-style, Slides → outline). Binary files like images returned with metadata only.
Best for: “find that contract from last March and pull the payment terms,” “read the latest quarterly review doc and extract the OKRs,” or any “I know it’s in Drive somewhere” research task that used to mean 10 minutes of search-clicking.
Skip if: your team is on Notion or Dropbox — wrong storage. Skip if your need is bidirectional (read + write) — this is read-only. For write, look at the broader Google Workspace MCP servers in the community.
Setup gotchas: OAuth setup, similar friction to Gmail MCP — Google Cloud Console, enable Drive API, create OAuth client, authorize. Server is in servers-archived (Anthropic moved active maintenance), but functional. Permission is per-Drive (yours by default); shared-drive access requires re-authorization with the right scope.
Real-world workflow: before any client call, the agent runs search_files "[Client Name]" newer_than:90d, reads the most-recent document with their name, summarizes the open items. Three minutes vs. 15 of “let me find that doc…” live on a call.
Compatible alternatives: Notion MCP if your knowledge actually lives in Notion, Filesystem MCP for local files instead of cloud.
Read-only is correct. If you ever want to overwrite a Drive Doc from an LLM, do it through a controlled “publish” flow, not a generic write tool.