Google Cloud API Gateway added support in Public Preview for serving existing REST operations as tools that AI agents can discover and call through the Model Context Protocol (MCP). Announced Sept. 24, 2026, the feature lets teams annotate an OpenAPI specification they already deploy and publish the operations through a remote MCP server, without building, hosting, or maintaining a separate server. Google says frameworks including the Agent Development Kit (ADK) and Gemini Enterprise support MCP natively.

API Gateway accepts MCP JSON-RPC requests at one endpoint, converts each tools/call request into its corresponding REST request, applies the operation’s existing policies, and returns the response as an MCP result. Existing JWT or API-key authentication, quotas, and logging continue to apply because MCP and REST requests share the same policy path; each operation uses one quota allocation regardless of how it is called. The gateway serves MCP under the /mcp base path after the API config is deployed, with no additional infrastructure to provision.

The OpenAPI document must use version 3.0.x or 3.1.x; OpenAPI 2.0 specifications are unsupported and must be migrated. Teams enable MCP for a document with x-google-api-management.mcp and can customize or exclude individual operations with x-google-mcp-tool. Each operation exposed as a tool needs a backend and a non-empty description. Google advises describing when and why an agent should use a tool, since that description guides the model’s decision to call it.

By default, tools/list discovery is unauthenticated, exposing tool names and input schemas to anyone who requests them. Production deployments can require a JWT for discovery; API keys cannot secure tools/list. Calls to tools/call always use the authentication required by the underlying REST operation. Developers can connect an MCP client to the gateway endpoint; Google’s ADK example supplies the endpoint and the gateway’s expected credential.

Connecting the gateway to API hub publishes its MCP server with MCP metadata and lists it automatically in Agent Registry. The Public Preview supports REST and OpenAPI 3.x backends with existing authentication. Google lists MCP resources and prompts, response streaming, and Model Armor payload inspection as roadmap items. Current limits include excluding operations that return empty bodies, such as HTTP 204; schemas with deeply nested objects may not display fully in tools/list; a gateway can serve up to 1,000 tools; and MCP and model routing cannot be enabled in the same API config.

Google Cloud API Gateway added support in Public Preview for serving existing REST operations as tools that AI agents can discover and call through the Model Context Protocol (MCP). Announced Sept. 24, 2026, the feature lets teams annotate an OpenAPI specification they already deploy and publish the operations through a remote MCP server, without building, hosting, or maintaining a separate server. Google says frameworks including the Agent Development Kit (ADK) and Gemini Enterprise support MCP natively.

API Gateway accepts MCP JSON-RPC requests at one endpoint, converts each tools/call request into its corresponding REST request, applies the operation’s existing policies, and returns the response as an MCP result. Existing JWT or API-key authentication, quotas, and logging continue to apply because MCP and REST requests share the same policy path; each operation uses one quota allocation regardless of how it is called. The gateway serves MCP under the /mcp base path after the API config is deployed, with no additional infrastructure to provision.

The OpenAPI document must use version 3.0.x or 3.1.x; OpenAPI 2.0 specifications are unsupported and must be migrated. Teams enable MCP for a document with x-google-api-management.mcp and can customize or exclude individual operations with x-google-mcp-tool. Each operation exposed as a tool needs a backend and a non-empty description. Google advises describing when and why an agent should use a tool, since that description guides the model’s decision to call it.

By default, tools/list discovery is unauthenticated, exposing tool names and input schemas to anyone who requests them. Production deployments can require a JWT for discovery; API keys cannot secure tools/list. Calls to tools/call always use the authentication required by the underlying REST operation. Developers can connect an MCP client to the gateway endpoint; Google’s ADK example supplies the endpoint and the gateway’s expected credential.

Connecting the gateway to API hub publishes its MCP server with MCP metadata and lists it automatically in Agent Registry. The Public Preview supports REST and OpenAPI 3.x backends with existing authentication. Google lists MCP resources and prompts, response streaming, and Model Armor payload inspection as roadmap items. Current limits include excluding operations that return empty bodies, such as HTTP 204; schemas with deeply nested objects may not display fully in tools/list; a gateway can serve up to 1,000 tools; and MCP and model routing cannot be enabled in the same API config.