Claude Code’s Code Mode: 62 Changelog Entries in 8 Days

Anthropic added an experimental Code Mode to Claude Code in version 2.1.323, published on 12 November, its changelog shows. Instead of asking for one tool call at a time, Claude writes a short JavaScript program that makes several calls in one step.

In the eight days to 20 November, the changelog mentions Code Mode in 62 entries across seven releases, by our count; 37 of them are fixes. It is off by default and, per Anthropic’s documentation, may change or be removed in any release.

Claude Code is Anthropic’s coding agent, one of the products behind the Anthropic stock story. Code Mode changes how that AI agent uses its tools, such as Bash to run commands, Read to open files and Edit to change them.

Business Pill · AI Agents

An evergreen Business Engineer explainer on AI agents, from answering to doing. It teaches the general idea only and does not discuss Claude Code.

The key insight: Code Mode lets Claude write one JavaScript program that makes several tool calls; it is experimental, off by default, and Claude Code’s changelog mentioned it 62 times in its first eight days, by our count.

What Does Code Mode Change?

Without Code Mode, Claude asks for a tool call, reads the result, then decides the next step. With it, Claude writes a piece of JavaScript that makes the tool calls, and Claude Code runs that code in the session, the docs say.

One piece of code can run commands in parallel, use one result in the next call and return only the part of the output Claude needs to read. Anthropic suggests it for changes that repeat across many files and for searches that need filtering.

In the conversation, users see the tool calls, such as an edit’s diff, and not the code. Pressing Ctrl+O opens the transcript viewer, where the code can be read.

By our count, Claude Code's changelog mentions Code Mode in 62 lines across seven releases from 12 to 20 Novem
By our count, Claude Code’s changelog mentions Code Mode in 62 lines across seven releases from 12 to 20 November, peaking at 22 in version 2.1.328.

Why Write Code Instead of Calling Tools?

The idea predates Claude Code. Cloudflare argued in September 2025 that language models are better at writing code to call tools than at calling them directly, and called its approach Code Mode.

Anthropic made a similar case in a November 2025 engineering post. In its example, having an agent write code to call tools cut token use from 150,000 to 2,000, a saving of 98.7%, Anthropic said.

Bar chart: 150,000 tokens calling tools directly vs 2,000 writing code, Anthropic's MCP example. Data: Anthropic
In Anthropic’s own November 2025 example, an agent that wrote code to call tools used 2,000 tokens instead of 150,000. Anthropic has not published a figure for Code Mode.

That example concerned tools connected through the Model Context Protocol, not Claude Code’s new mode, and Anthropic has not published a comparable figure for Code Mode.

Who Can Turn It On?

The requirements are narrow for now. Users need version 2.1.323 or later, a Claude subscription or Console account, an interactive session in a terminal, and Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable 5.2.

It does not work through Amazon Bedrock, Google Cloud, Microsoft Foundry or a gateway, nor for organisations on the HIPAA or zero-data-retention setups. The VS Code panel, the desktop app and cloud sessions are excluded too.

Developers turn it on with the environment variable CLAUDE_CODE_EXPERIMENTAL_CODE_MODE set to 1. Claude Code ignores that setting inside a project’s files, so a cloned repository cannot switch it on, and administrators can block it company-wide.

What Are the Guardrails?

Each tool call inside the code still passes through the user’s permission rules and hooks under its usual name. In auto mode, though, the safety classifier reviews each piece of code before it runs, not each call, the docs say.

The docs add that the JavaScript itself and its file, web and MCP tool calls run outside the sandbox, while shell commands follow the user’s sandbox settings. Code that waits stops after ten minutes; code in the background can run for up to two hours.

How Fast Is It Changing?

Quickly. Release 2.1.328 on 18 November alone carried 22 Code Mode entries, about 35% of the 62, our arithmetic. One later fix stopped “a runaway loop in Code Mode filling the machine’s memory.”

We covered a different product with a similar name in October, when Amp opened its Claude Code Mode to Claude Pro and Max plans. Anthropic’s feature is built into Claude Code itself.

The Structural Read

As we read it, Anthropic is moving Claude Code toward the approach its own engineers and Cloudflare tested on outside tools: fewer model round trips per task.

As we read it, the narrow requirements and fast changelog suggest Anthropic is still testing Code Mode with a limited group before any wider default.

Claude Code documentation

“With Code Mode, Claude writes a piece of JavaScript that makes the tool calls, and Claude Code runs that code in your session.”

Three Implications

FOR DEVELOPERS Code Mode is opt-in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1, needs version 2.1.323 or later and works only in an interactive terminal session.

FOR SECURITY TEAMS In auto mode the classifier reviews each piece of code rather than each call, and the JavaScript and its file, web and MCP calls run outside the sandbox; administrators can block it with managed settings.

FOR COST WATCHERS Anthropic’s earlier MCP example cut tokens from 150,000 to 2,000, but it has published no Code Mode figure.

What Is Not Established

Anthropic has not said how much Code Mode saves in tokens, time or cost, or how many developers use it. Our count of 62 entries reflects changelog lines that mention Code Mode, not the size of each change.

This Week In The AI Supercycle

A World Without an Order, and an AI Central Bank Without a Successor

This week’s issue: the week AI memory stayed largely sold out while the people holding the shares paid: why the AI Supercycle is not a dot-com replay, read through geopolitics, macro and technology, with the gap between building and financing the widest we’ve recorded.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of who builds, finances and captures value in AI.

Read the Map of AI →

The Bottom Line

Code Mode lets Claude batch several tool calls into one script, an idea Cloudflare and Anthropic first tested on outside tools. It is experimental and narrow for now, and it changed 62 times in its first eight days, by our count.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

Frequently Asked Questions

What is Code Mode in Claude Code?

Code Mode is an experimental Claude Code feature in which Claude writes JavaScript that makes several tool calls, such as Bash, Read and Edit, in one step, and Claude Code runs that code in the session. It is off by default, Anthropic's documentation says.

How do you turn on Code Mode?

Set the environment variable CLAUDE_CODE_EXPERIMENTAL_CODE_MODE to 1 in your shell or user settings, on Claude Code version 2.1.323 or later. It works only in an interactive terminal session with Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable 5.2.

Does Code Mode save tokens?

Anthropic has not published a figure for Code Mode. In a November 2025 post on code execution with MCP, Anthropic said writing code to call tools cut token use from 150,000 to 2,000 in its example, a 98.7% saving.

Can companies block Code Mode?

Yes. Administrators can set CLAUDE_CODE_EXPERIMENTAL_CODE_MODE to 0 in managed settings so nobody in the organisation can use it, the docs say. Project settings files cannot turn it on.

A note on sourcing. How Code Mode works, its requirements and its limits come from Anthropic’s Claude Code documentation. Release contents come from Claude Code’s public changelog, and release dates from the npm registry’s publish times. The 150,000-to-2,000-token example and the 98.7% figure are Anthropic’s, from its November 2025 post on code execution with MCP. The 62-entry count, the per-release counts, the eight days and the 35% share are our arithmetic from the changelog.

Sources: Claude Code docs, ‘Run tool calls from code with Code Mode’ (read 22 Nov 2026) · Claude Code CHANGELOG, versions 2.1.323 to 2.1.330 (npm publish dates 12 to 20 Nov 2026) · Anthropic engineering, ‘Code execution with MCP’ (4 Nov 2025) · Cloudflare, ‘Code Mode: the better way to use MCP’ (26 Sep 2025) · Our 248346: Amp opens its Claude Code Mode to Claude Pro and Max plans

Scroll to Top
FourWeekMBA