Claude Code 2.1.323, released on 12 November 2026, adds an experimental, opt-in Code Mode in which Claude writes JavaScript to call its tools, and it ships alongside dozens of Bash permission and sandbox fixes. Our verdict: standard tool calls win for day-to-day work today, and the upgrade is worth taking for the hardening alone. Trial Code Mode only in a sandboxed or throwaway project until its documentation is live and your team has decided who may turn it on.
TL;DR
- Claude Code 2.1.323 adds Code Mode, an experimental opt-in where Claude writes JavaScript to call its tools. Organisations can turn it off.
- The same release fixes dozens of Bash permission and sandbox problems, several of which let commands run without a prompt.
- Verdict: keep standard tool calls as your default and upgrade for the fixes. Treat Code Mode as a supervised experiment.
- The Code Mode documentation page returned a 404 when we checked on 13 November 2026, so this article does not include setup steps.
- If you use plugins, the new
engines["claude-code"]field inplugin.jsonstops plugins loading outside their declared version range.
What is Code Mode in Claude Code 2.1.323?
The official changelog describes it as an experimental opt-in in which Claude writes JavaScript to call its tools. The entry points to a docs page for turning it on, and for turning it off across an organisation.
Under standard tool calling, the model emits one structured request per tool call, waits for the result, then decides what to do next. Code Mode changes the unit of work. Claude writes a short program that calls several tools, and the program handles the intermediate results itself. Amazon Web Services' partner blog published a walkthrough of the release the same morning and flagged the same org-level off switch as the notable governance detail.
A concrete example: finding every file that mentions a deprecated function, then reading the three largest. With standard calls that is a search, a result returned to the model, then three reads, each a round trip. In Code Mode it could be one script that searches, sorts and reads in a single execution.
Code Mode vs standard tool calls: which should you use?
The design intent is fewer round trips through the model, which should reduce overhead in long sessions. We have no measurements of that for 2.1.323, and Anthropic has not published any that we could find, so treat it as a hypothesis.
| Standard tool calls | Code Mode | |
|---|---|---|
| Status | Default, mature | Experimental, opt-in |
| What you review | One discrete call at a time | A script that may make many calls |
| Organisation control | Existing permission rules | Can be turned off organisation-wide |
| Documentation | Available | Docs page returned 404 on 13 Nov 2026 |
| Best for | Day-to-day edits, reviews, anything on a real repository | Supervised trials in disposable projects |
Standard tool calls win for now because their failure modes are known. A permission prompt shows you one action. A script can bundle several actions into a single approval, so reviewing it takes more care. That is a tradeoff, not a flaw, but it matters when the surrounding permission system has just needed this many fixes.
What does Claude Code 2.1.323 change for Bash permissions?
Most of this release is not Code Mode. It is a hardening pass on the places where a prompt-injected instruction could slip through. The changelog lists these among its fixes:
- Some
lsofcommands ran without a permission prompt. - The Bash sandbox could switch off during a session.
- Allow rules with an attached wildcard, such as
Bash(ls*), skipped the prompt in some cases. A rule that names a variable, such asBash(NODE_ENV*), now asks. - Permission checks could judge an
rmfrom the wrong folder and skip the dangerous-removal prompt under bypass permissions or a Bash allow rule. - Sandboxed commands could read Claude Code's file checkpoint copies under the config directory.
-
claude -pand the Agent SDK ran Bash unsandboxed when--settingsor SDK settings that enable the sandbox failed validation. Claude Code now stops at start-up and names the invalid setting.
A further fix concerns auto mode. If --permission-mode auto or defaultMode: auto is set but auto mode is unavailable, Claude Code previously started in a mode that could allow more than intended. It now starts in Manual mode, which asks first. The release notes index confirms the same security batch, including the wildcard allow-rule fix and the new engine identifier for plugins.
The wildcard fix is the one to act on. If your settings contain allow rules such as Bash(ls*), review them after upgrading. Some will now prompt where they previously did not, and that is the correct behaviour.
How does this release treat plugins?
Three changes affect plugin authors and anyone installing third-party plugins:
-
plugin.jsonacceptsengines["claude-code"]. Plugins outside their stated Claude Code version range are not installed, updated or loaded, and install and update choose the newest version that fits. -
claude plugin validatenow type-checks a plugin's TypeScript hooks module. Each finding is a warning, and--strictmakes it fail. - Mods can no longer reword Claude Code's own confirmation questions.
The last item is worth noting for security. A confirmation prompt is only useful if the text is trustworthy, so stopping extensions from rewriting it removes a real social-engineering route.
Should you upgrade to Claude Code 2.1.323?
Yes, for the fixes. The 2.1.323 release arrived two days after 2.1.321 and three days after 2.1.322, and the Havoptic release tracker counts it as the 262nd Claude Code release of 2026. Independent version trackers mirror the same changelog batch. At that cadence, pinning a version for months is hard to justify when the notes describe prompts being skipped.
A sensible sequence:
- Upgrade a single machine and run your usual workflow with standard tool calls.
- Review any Bash allow rules ending in a wildcard.
- If you use
claude -por the Agent SDK in CI, confirm your sandbox settings validate, since an invalid setting now stops start-up instead of running unsandboxed. - Decide at organisation level whether Code Mode should be available at all, before anyone enables it individually.
Secrets handling matters more once a model writes executable code. Our guide to Bitwarden vs 1Password for Claude Code secrets covers keeping credentials out of reach. For the previous release's default-model and pricing changes, see our Claude Code 2.1.321 coverage. If you are weighing how much to trust agent output, verify-first vs build-first is a useful frame, and our piece on subagent cost and status updates covers the spending side.
What does it cost, including in India?
Claude Code is included with Claude Pro and Max subscriptions, and Anthropic lists prices on its pricing page. Pricing is in US dollars. Indian users generally pay by international card, and GST may apply to imported software services, so check the final charge at checkout. This release does not change the pricing structure.
Why we cover this
We choose topics partly by measuring demand. In our own keyword corpus (n=695, measured 12 November 2026), we priced 695 keywords in the AI and developer-tooling space using DataForSEO volume and difficulty data. Only a small share cleared our bar for a winnable topic: modest search volume, low difficulty, a genuine technical term, and a phrase of at least three words. Version-specific release coverage like this is a narrow niche, and the audience for it is people who need to decide what to do this week.
FAQ
Q: What is Code Mode in Claude Code 2.1.323?
A: It is an experimental, opt-in feature in which Claude writes JavaScript to call its tools instead of issuing each tool call separately. The changelog says organisations can turn it off.
Q: How do I enable Code Mode?
A: The changelog points to a documentation page for this, but that page returned a 404 when we checked on 13 November 2026. We have not verified the steps, so check the official docs before relying on any guide.
Q: Is Claude Code 2.1.323 safe to upgrade to?
A: The release is mostly fixes, including several that close cases where Bash commands ran without a prompt. Upgrade one machine first and review any wildcard allow rules, since some will now ask where they did not before.
Q: Does Code Mode make Claude Code cheaper or faster?
A: The idea is fewer model round trips in long sessions, but we found no published measurements for 2.1.323. Treat any savings as unproven until you test on your own workload.
Q: What happens if auto mode is unavailable?
A: Claude Code now starts in Manual mode, which asks before acting. Previously it could start in a mode that allowed more than intended.
Q: Can plugins break after this update?
A: Plugins that declare an engines["claude-code"] range are not installed, updated or loaded outside it. Authors should run claude plugin validate, optionally with --strict, to catch hooks type errors.
Top comments (0)