Workers AI, AI Gateway, AI Search, and MCP

Try a Workers AI model#

Enable Workers AI in the account’s active profile. Open its model catalogue, filter by task/model, and inspect a model’s overview before opening its playground. Text, translation, embeddings, and image tasks have different input shapes; use the fields generated for the selected model.

Submit a small identifiable request, inspect the output, and check usage/errors. A model appearing in a catalogue is not a guarantee that the account can invoke it. The provider’s service limits and permissions still apply.

Use AI Gateway#

Open a gateway’s Chat Playground, choose provider and model, and supply provider credentials when that provider requires them. Streaming and non-streaming responses serve different inspection needs. Skip cache changes the test’s cache behavior; it is useful when comparing a new request with an earlier cached response.

In Logs, combine text, provider, success/error, and cache filters, then refresh. Analytics summarizes the loaded log set; a low total in a narrow view is not necessarily the gateway’s lifetime usage. The model catalogue’s Try it selects a model in the playground rather than deploying a model.

Inspect AI Search#

Use AI Search for index/retrieval-oriented work. Inspect the selected index and indexing state before trying retrieval or chat. Search results depend on the indexed content and current state; an empty result can come from indexing, filters, or the question rather than a failed chat model.

Connect an MCP server#

Configure enabled MCP servers in Cloudflare settings, enable the category in the profile, and refresh the explorer. Open a server dashboard and Connect, then Refresh tools. Inspect each tool’s schema before entering arguments and executing it.

A simple tool argument body might look like:

{"query":"example"}

Use the server’s actual schema; query is not universal. Review results and success/error history. Disconnect clears the active connection/session state; reconnect and reload tools after server changes. MCP tool execution can perform whatever operation that tool implements, so inspect its target and arguments as carefully as a direct service action.