Google added four capabilities to the Gemini API's Managed Agents: remote MCP (mcp_server), background execution, function calling with a transition to requires_action, and credential refresh via environment_id without losing environment state; the least flashy piece — renewing the token without restarting — is what makes a long-running agent viable.
Some announcements are read for what they say, and others for what they let slip. The July 7 one belongs to the second group. Google — through the pens of Philipp Schmid and Mariano Cocirio, of Google DeepMind — introduced four new capabilities for the Gemini API's Managed Agents, and one of them is support for remote MCP servers. It is not just another marketing checkbox: it is a plumbing decision. When a frontier lab decides that its agents will talk to external tools through the Model Context Protocol, it is taking a side on what the default wiring of this generation of software will be. And plumbing decisions, precisely because they are invisible, are the ones that most shape what can be built on top.
What Google announced exactly
Four pieces, and it is worth looking at them one by one, because together they tell a story. The first is background execution: you pass `background: true`, the server runs the interaction asynchronously and immediately hands you back an ID to poll for status. The second is integration with remote MCP servers via an `mcp_server` tool, which lets the agent reach private databases and internal APIs without custom middleware. The third is custom function calling: the built-in tools run on their own on the server, while your own functions push the interaction into `requires_action` so your client can run the local business logic. And the fourth — the quietest — is credential refresh: you renew the token by passing your existing `environment_id` with a new network configuration, without losing the filesystem state or the packages already installed.
Why the boring detail is the one that matters most
Of the four, the one that will let engineers breathe is the last, and it is the least flashy. Anyone who has built a long-running agent knows the drama: the access token expires halfway through a twenty-minute task, and the naive response is to restart the whole environment, reinstall dependencies and throw away everything the agent had already done. Refreshing the credential while keeping the state is the difference between an agent that dies every hour and one that can actually work for hours. It is changing the flashlight's battery without turning it off: little glamour, huge consequence. Placed right next to background execution, the intent is clear: agents that run on their own, for long stretches, touching real systems.
MCP as wiring, not as a lock
Here is where it pays to separate the fact from the reading. The fact is that Google brought remote MCP into its agents API. The reading — broader — is that MCP, the open protocol Anthropic released and that has been getting adopted across the ecosystem, is becoming the standard substrate the rest is built on, the way HTTP or USB once were. Be careful not to confuse the layers: MCP is not the lock that decides who gets into your data — that is still handled by OAuth and API-Keys — but the standardized pipe through which the conversation between the agent and the tool flows. A frontier lab taking it for granted does not prove a declared convergence of the whole industry, but it is a strong signal: pipes, once they start to standardize, tend to become invisible and everywhere.
What it means for people building with AI
The lesson for anyone shipping products with agents is direct: always tell the lock apart from the wiring, and do not tie your architecture to this quarter's fashionable lock when what endures is the pipe. That same reasoning — MCP understood as a transport standard, not as a permission mechanism — is what guides our own reading at NeuralOS, and an announcement like Google's works as external validation of that hypothesis. There is a second honest nod: the pattern of refreshing credentials without restarting the environment is exactly the problem that a vault of per-account encrypted keys exists to solve, renewing tokens without breaking what the agent already had running. When a big lab ends up building a piece you already considered necessary, it does not mean you arrived late: it means you were in the right lane. And that, for anyone who builds, is the best place to be.