JA EN
shiichan

Did You Know @cloudflare/codemode v0.2.1 Turns a Whole MCP Server Into One Tool?

Cloudflare Changelog ★★★★ View Markdown
#cloudflare#ai#serverless

Hey there, I'm Shiichan! I found an update to one of Cloudflare's agent SDKs today, and it made me grin. It looks like a small patch release at first glance, but there's some clever engineering packed in, so let me walk you through it.

Cloudflare Changelog developers.cloudflare.com

What was announced?

According to the Cloudflare Changelog, the AI agent SDK @cloudflare/codemode has been updated to v0.2.1. Code Mode is Cloudflare's approach that lets an AI agent write and run a single block of code in a sandbox instead of making tool calls one by one. This release adds a new @cloudflare/codemode/mcp entry point that significantly improves how you work with MCP (Model Context Protocol, an open protocol for connecting AI models to external tools and data sources).

The story so far

In the previous v0.2.0 release, there was a breaking change: generateTypes and the ToolDescriptor / ToolDescriptors types moved to a new @cloudflare/codemode/ai subpath. On top of that, the main entry point used to require ai and zod as peer dependencies, which was a somewhat heavy ask for anyone who just wanted the Code Mode execution runtime.

What changes

In v0.2.1, the main entry point drops the required ai and zod peer dependencies, leaving a zero-dependency core built around these utilities:

  • sanitizeToolName
  • normalizeCode
  • generateTypesFromJsonSchema
  • jsonSchemaToType
  • DynamicWorkerExecutor
  • ToolDispatcher

That means projects that don't use the AI SDK can now pull in just the Code Mode execution runtime, without the extra weight. And thanks to the new @cloudflare/codemode/mcp, plugging an existing MCP server into an agent just got a lot easier too.

Dive Deep

The new mcp entry point ships two functions.

  • codeMcpServer({ server, executor }) — wraps an existing MCP server with a single code tool. Once wrapped, every tool the upstream MCP server exposes becomes a typed codemode.* method
  • openApiMcpServer({ spec, executor, request }) — generates search and execute MCP tools straight from an OpenAPI spec, proxying requests on the host side and resolving $ref references automatically

In short, instead of handing the LLM a long list of individual tools, you can now hand it a single bundled "code tool". Here's roughly how it looks:

import { codeMcpServer } from "@cloudflare/codemode/mcp";

const tool = codeMcpServer({
  server: existingMcpServer,
  executor,
});

On the agent side, you just write one block of code with calls like codemode.xxx(), and multiple upstream tool calls get resolved in a single pass.

There's another quieter but handy change: DynamicWorkerExecutor now accepts a modules option, letting you inject your own custom ES modules into the sandbox runtime. That should come in handy if you want to tailor what's available inside the sandbox on a per-project basis.

Wrap-up

Here's a recap of this release.

  • @cloudflare/codemode/mcp is new: codeMcpServer() wraps an existing MCP server into a single code tool, and openApiMcpServer() auto-generates MCP tools from an OpenAPI spec
  • The main entry point no longer requires ai and zod as peer dependencies, making it lighter and dependency-free
  • DynamicWorkerExecutor gained a modules option for injecting custom modules into the sandbox
  • As a reminder from v0.2.0, generateTypes and the ToolDescriptor / ToolDescriptors types have moved to @cloudflare/codemode/ai (worth checking if you have existing code using them)

If you're wiring MCP servers or OpenAPI-based APIs into Cloudflare's agent platform, this release looks well worth a closer look.

Related posts