/
Published on 12.03.2026
TLDR: Google quietly released a Workspace CLI built explicitly for AI agents, and when combined with Anthropic's new /loop command in Claude Code, it creates a system where your AI assistant can autonomously manage email, calendar, docs, slides, and drive on a recurring schedule throughout your workday. The result is an "agentic workspace" that removes the human as the middleman between AI tools and the communication layer where real work happens.
There is a newsletter making the rounds that walks through what happens when you wire together Claude Code, Google Workspace CLI, Obsidian, and a handful of MCP integrations into a single daily operating system. The author has been building this incrementally over months, starting with consolidating writing and content workflows into Claude Code, then layering in research via NotebookLM MCP, image generation through Nano Banana, web research via Perplexity and Firecrawl, and most recently a full project management system inside Obsidian with slash commands for planning, processing, and reviewing work.
The persistent gap was Google Workspace. Email, calendar, documents, slides, and sheets remained trapped in browser tabs, requiring the human to act as the bridge between Claude and the communication tools where external collaboration actually happens. Google's new GWS CLI closes that gap by giving the terminal direct API access to Gmail, Calendar, Docs, Drive, Sheets, and Slides. The key distinction the author emphasizes is that this is not an MCP server maintained by a third party. Google built and maintains the CLI themselves, which means when their APIs change, the tool updates accordingly. No waiting for community maintainers to patch a broken bridge. Claude runs the gws command natively in the terminal, the same surface where it already operates.
The practical daily workflow described is genuinely interesting. A morning briefing fires automatically when the session starts, categorizing dozens of emails into segments, summarizing newsletters, flagging items for reply, and pulling up the day's calendar. Before a client call, a single command pulls the attendee's email history, finds shared documents in Drive, and compiles a one-page meeting prep. After the call, a transcript gets processed into a summary with action items routed to the Obsidian vault, tagged with deadlines and linked to the right sprint. Draft follow-up emails get generated for review. An end-of-day inbox zero command categorizes everything, drafts replies, routes action items, and archives the rest. The author reports this takes ten minutes instead of forty-five.
What makes this architecturally notable is the convergence of two separate releases. Google provided the data access surface, and Anthropic shipped the /loop command that lets Claude Code execute workflows on recurring intervals within an active session. Every 10, 30, or 60 minutes, Claude can check for upcoming meetings, process new emails, or run whatever scheduled task you have configured. This is not a cron job or external automation. It runs inside your live Claude Code session for as long as the session is open. The distinction between CLI and MCP is worth thinking about carefully. MCP is a protocol that requires a translation layer between Claude and external tools, maintained by whoever built the server. A CLI is the tool itself, speaking the terminal's native language. Both have their place, but the reliability argument for first-party CLIs is strong. The author's MCP integrations for NotebookLM and Nano Banana work well but depend on community maintenance. The GWS CLI has Google behind it.
There is a legitimate question about how much autonomy to grant here. The author addresses this directly with permission whitelisting: Claude can read your inbox, check your calendar, and search Drive without asking, but still requires explicit approval before sending anything on your behalf. That is a sensible boundary, though anyone setting this up should think carefully about what "read access to your entire inbox" means in practice. The convenience is real, but so is the attack surface if your Claude Code session is ever compromised. The broader thesis, that every software company needs to ask whether their product is easy for agents to access, echoes a point Andrej Karpathy has been making. The companies building CLI and terminal-native interfaces for their products are positioning themselves for a world where AI agents are the primary users of software, not humans clicking through browser UIs.
How Google Workspace CLI Made My Claude Code Setup 10x More Powerful