ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
Claude Code's New Code Mode: Letting the Model Script Multiple Tool Calls at Once

Claude Code's New Code Mode: Letting the Model Script Multiple Tool Calls at Once

AI information • Admin • • 27 views

Claude Code introduced its experimental Code Mode on November 17: instead of letting Claude call one tool at a time, it has the model write a piece of JavaScript that Claude Code runs directly in the session, completing several tool calls in one step. The change shipped with version 2.1.323 on November 12 and is off by default — it has to be turned on by hand.

From "One Call at a Time" to "One Script Does It All"

Until now, Claude Code worked as "call a tool, read the result, decide the next step" — the more tools a job needs, the more round trips pile up. Code Mode flips that around: Claude writes a piece of code that calls the tools first, Claude Code executes it inside the session, and one script can run several commands in parallel, pass one result into the next call, and return only the part of the output Claude needs to read. Anthropic recommends it for jobs with many related tool calls, such as making the same change across many files or searching and then filtering the results. The feature is switched on with the environment variable CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1, and organization admins can disable it through managed settings. The Claude Code team has previously disclosed the validation method behind its claim that ninety percent of its code is written by AI without errors.

The experimental label comes with plenty of limits: it only works in interactive terminal sessions, so the Desktop app, the VS Code panel, -p mode, the Agent SDK, and cloud sessions are all out; models are limited to Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.2; and cloud gateways such as Bedrock, Google Cloud, and Foundry are not supported either. In auto mode, the reviewer checks the whole script rather than each individual tool call, and it cannot see values computed at run time; code that is waited on is cut off after ten minutes. Anthropic has published no figure for how many tokens Code Mode saves — the widely shared 37% comes from an article about programmatic tool calling in the API from last November, not from measurements of this feature, so don't treat it as an official promise.

Start With Repetitive Jobs; Don't Rush Production Workflows Onto It

Anyone already on 2.1.323 or later can try it in a terminal, best on repetitive jobs like "change the same line in a dozen files"; production workflows and critical operations involving external services should wait. Version 2.1.325 already fixed several early bugs, so update before trying it out. Whether Code Mode graduates from experimental depends on whether it proves stable and actually saves tokens over the next few releases — until then, it is more of a switch for heavy users than a default everyone should flip on.

Recommended Tools

More