Anthropic has rolled out Code Mode for Claude Code — an experimental, opt-in feature that lets the CLI agent write and execute JavaScript to chain multiple tool calls in a single step. It shipped with Claude Code v2.1.323 last week and is now live for developers using the CLI on macOS, Linux, or Windows via WSL.
No web or desktop app version supports it yet. The feature is not enabled by default: users must set CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 in their shell or in ~/.claude/settings.json, then confirm the prompt at session start. Once on, it shows as ‘Code Mode (experimental) is on’. Flip the variable to 0, and it’s off, even if another config file tries to turn it back on.
Enterprise admins can block it globally but cannot enforce it across teams. Code Mode changes how Claude handles multi-step workflows. Instead of reading a file, then editing it, then running tests, each requiring a separate round-trip, Claude now writes a short JavaScript snippet that does all three in sequence or parallel.
One documented example finds a file, edits it, and runs tests, with retries built in. You won’t see the code in the chat unless you press Ctrl+O to open the transcript viewer and expand a collapsed Code row. All tool calls inside the script still respect your deny/ask rules and run PreToolUse hooks under their usual names (e.g., Bash).
Hook authors need to parse commands carefully: a git push might arrive as cd 'app' && git push. To inspect the generated code itself, match the REPL tool and read tool_input.code. Execution time limits apply. If Claude waits for permission, the script times out after 10 minutes, including time spent at prompts.
In the background, while you keep chatting, it can run up to 2 hours, but only if the session started in auto or bypassPermissions mode. Anthropic doesn’t publish official token savings for Code Mode in the CLI, but one third-party test reported system-tool context dropping from 24.6k tokens to 11.1k tokens when enabled.
That’s consistent with earlier Anthropic benchmarks using Programmatic Tool Calling in Python, where complex research tasks cut average token use from 43,588 to 27,297. Cloudflare coined the term ‘Code Mode’ in September 2025 for its Agents SDK. Anthropic adopted the name after publishing its own MCP-based code execution work. Documentation is live.
So is the changelog. No price change, no new subscription tier, just a new way to call tools faster, with tighter control over flow and context. Developers who rely on Claude Code for local dev loops, CI scripting, or repo triage now have a lower-latency path through multi-step tasks. The feature remains experimental.
Rollout is global. No regional restrictions. No Indian or US pricing to report, it’s free for all existing Claude Code users. Next up: wider tool support, better error visibility, and possibly language-agnostic code generation, but Anthropic hasn’t confirmed any of that.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.