Give Gemini the web, on your terms.
Gemini can't call an API on its own. AlwaysOn MCP is a gateway it connects to: Gemini asks, the gateway makes the real GET or POST, and only to the hosts you allow.
Works with Gemini's custom MCP connectors. Setup takes a few minutes.
Three steps. Then Gemini can make the call.
Your account, your allowlist, your connection. Nothing is shared with other users.
List the hosts you allow.
Sign in with Google and add hosts: an exact one, a whole domain with *.example.com, or every public site with *. Tick GET, POST or both.
Your own client ID and secret.
One click makes a client ID and secret tied to your account. The secret is shown once and stored encrypted.
Paste it into Gemini.
Add the MCP URL in Gemini's connector settings, paste your client ID and secret, and approve. Every call now runs as you.
Open to the hosts you choose. Closed to everything else.
A gateway that makes web requests for an AI needs hard limits. These are on for every account.
Allowlist per user
Each account has its own list of hosts and methods. Edit it any time; changes apply to the next call.
Keys stay on the server
API keys are added by the gateway for the host they belong to. Gemini never sees them, and they never follow a redirect to another site.
Internal networks blocked
Private, loopback and cloud-metadata addresses are refused, including through redirects, even when you allow every host.
Every call knows who you are
Your connection is tied to your Google account, never to what the model types. Ask Gemini to run whoami to check.
Revoke in one click
Delete a client ID and its Gemini connection stops on the very next call. No waiting for tokens to expire.
Sensible limits
Timeouts, response size caps and per-account rate limits keep a runaway conversation from turning into a runaway bill.
A standard MCP server, not a walled garden.
Three tools over Streamable HTTP with OAuth. Any MCP client that supports OAuth can connect the same way Gemini does.
http_getandhttp_post, checked against your allowlistwhoamireturns the account and allowlist the gateway sees- Responses come back with status, headers and body, marked as untrusted content
# MCP server URL https://mcp.signaltradeinc.com/mcp # Tool call from Gemini { "name": "http_post", "arguments": { "url": "https://api.example.com/v1/items", "json_body": { "name": "Weekly report" } } } # Gateway response { "status": 201, "content_type": "application/json", "body": "…" }
A few things to know.
Why can't Gemini make API calls by itself?
Gemini runs in Google's environment and doesn't send arbitrary requests to the internet. It can call tools on an MCP server you connect. AlwaysOn MCP is that server: Gemini asks for a request, and the gateway makes it.
Can Gemini reach any website?
Only the hosts on your allowlist, with the methods you tick. You can allow everything with *, but internal and cloud-metadata addresses stay blocked either way.
Where do my API keys go?
Into the gateway's server-side settings, referenced by name in your allowlist. The gateway adds the key only to requests for that host. Gemini never sees it.
What happens if a client secret leaks?
Connecting also needs your Google sign-in and an approval on the consent page, so the secret alone isn't enough. Delete the client ID to cut off its connection immediately, then create a new one.
Does it work with clients other than Gemini?
Yes. It's a standard MCP server over Streamable HTTP with OAuth, so any MCP client that supports a client ID and secret can connect.