You picked your AI agent. It writes code, drafts emails, summarizes reports. Then you ask it to check your Gmail, and it stares back like a cat locked outside a glass door. Three billion people live in Google Workspace — and for agents not named Gemini, that data sat behind a wall.

Google Widens the Pipe

On April 17, Google expanded official MCP support to Maps, BigQuery, and a growing list of cloud databases. This follows the gws CLI release six weeks earlier — an open-source MCP server that hit 10,000 GitHub stars in week one and covers Gmail, Drive, Calendar, Docs, Sheets, Slides, Chat, and Admin. Over 100 pre-built skills. Impressive README.

MCP — Model Context Protocol — acts as a universal plug standard for AI agents. Think USB, but for data. Run gws mcp, and Claude Desktop, VS Code, or any MCP-compatible agent can theoretically reach your entire Google universe.

I say "theoretically" because I actually wired Claude Desktop to gws. And "open" undersells the friction by about three OAuth screens and a prayer.

The Part the Blog Post Skipped

Every Google API scope demands explicit user approval. Want your agent reading Gmail AND writing Docs AND checking Calendar? Three separate permission grants, each behind Google's increasingly paranoid verification flow. For unverified apps — which yours will be unless you enjoy a 4-6 week review process — users see a full-screen warning that essentially reads "this app will harvest your organs."

The --sanitize flag pipes prompts through Google Cloud Model Armor to block prompt injection (when someone hides malicious instructions inside data your agent reads). Smart move. Also adds 200-400ms latency per call. Your agent now reads email at the speed of cuneiform translation.

Then rate limits hit. Gmail API allocates 250 quota units per user per second — sounds generous until your agent searches, reads, and summarizes a single thread. Three calls minimum, complex queries burning units fast. The new BigQuery MCP endpoints? 100 concurrent queries per project by default. Your agent hits that ceiling the moment it gets curious about your data warehouse.

Meanwhile, Gemini Lives Inside the House

Google's own agents, built through Workspace Studio (GA since March 19), skip all of this. No OAuth dance. No consent screens. No rate limit sweat. They render natively inside Gmail's sidebar, inside Docs, inside Calendar — not as an external app shouting through an API pipe, but as a built-in UI element operating on internal data representations.

Third-party agents serialize your data into JSON, push it through HTTP, parse the response, pray the context window holds. Gemini agents skip the serialization entirely. Fewer hops. Lower latency. Richer context.

Google handed everyone a front door key. Gemini takes the service elevator. Both reach the same floors. One doesn't wait for the bouncer.

Three Questions for Vegas

Google Cloud Next runs April 22-24. The keynote promises "The Agentic Enterprise" sessions and third-party agent governance talks. Three questions determine whether the MCP expansion matters or just decorates a press release:

Rate limit parity. Do third-party MCP agents get the same quota as Gemini-native agents? Nobody outside Google knows Gemini's internal rate limits. If they're higher — and they almost certainly are — that's competitive advantage baked into plumbing, not brains.

OAuth for headless agents. The current consent flow assumes a human clicks "Allow" once. Autonomous agents — like Anthropic's Managed Agents, launched April 8 with Notion and Asana but conspicuously without Google Workspace — need a completely different auth model. Will Google build one, or let that gap serve as Gemini's velvet rope?

Write access depth. The April 17 expansion opened Maps and BigQuery, but the MCP endpoints lean read-heavy. Can a third-party agent create Calendar events from BigQuery results? Update Sheets from Maps routing data? The answer separates a window from a door.

The Uncomfortable Math

The productivity APIs your agent can reach — and how fast — set its usefulness ceiling. Not the model's IQ. Google opened the MCP spec to everyone, and that's genuinely useful. But opening a spec and leveling a playing field are different operations, and Google has deep institutional memory for the difference.

Your agent can see your Google data now. Whether it can touch that data as quickly as Google's own — that's what Vegas needs to answer this week.