Code Mode

Traditional MCP exposes every endpoint as a tool:

• A GitHub MCP server exposes 50+ tools (list issues, create PR, search code…)
• Add Stripe, Slack, and your internal API — now you're at 200+ tools
• Every tool's schema (name, description, parameters) is injected into the LLM's context window on every request
• The LLM was fine-tuned on synthetic tool-call examples — not real-world usage

Step 1 — MCP schema → TypeScript API

Cloudflare's Agents SDK reads each MCP server's tool schemas and generates typed TypeScript interfaces. The LLM sees code, not tool definitions.

Step 2 — One tool, not dozens

The agent exposes exactly one tool to the LLM: codemode({code: string}). The LLM writes TypeScript, and the runtime executes it in a sandboxed V8 isolate.

Step 3 — Sandboxed execution

Code runs in a fresh Cloudflare Worker isolate (not a container). Starts in milliseconds, costs almost nothing, and is thrown away after each run. The isolate has zero internet access — it can only reach MCP servers through typed bindings.

Step 4 — Results flow back

The sandboxed code uses console.log() to return results. Only the final output enters the LLM's context — not the intermediate API responses.

1. Install

npm install @cloudflare/codemode

2. Wrap your tools with codemode

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

const { system, tools } = codemode({
  system: "You are a helpful assistant",
  tools: { /* your MCP servers */ },
});

3. Use in your agent

const stream = streamText({
  model: openai("gpt-5"),
  system,
  tools,  // ← just one tool: codemode({code})
  messages: [{
    role: "user",
    content: "Deploy a new Worker named hello-world"
  }]
});