Skip to content

Claude Code 2.1.323 adds Code Mode and plugin version ranges

Updated Claude Code

Summary

Claude Code 2.1.323 (Nov 12, 2026) adds experimental Code Mode, engines ranges for plugins and a 1M default context for unknown models, plus dozens of permission and sandbox fixes.

Date
MCP
Yes
Pricing
Not stated in the sources we cited.

Claude Code v2.1.323 reached npm on November 12, 2026 at 20:29 UTC (23:29 Istanbul time) and is the current latest tag. It is the first build we cover since 2.1.316, and the notes are long: about 290 entries, roughly 150 of them fixes. The headline is Code Mode, an experimental opt-in where Claude writes JavaScript to call its tools. Most of the rest tightens permissions and the Bash sandbox, and a few defaults change in ways that can surprise teams on gateways or other vendors' models.

Release facts

FactDetail
ProductClaude Code, Anthropic's terminal coding agent
Version2.1.323
Release dateNovember 12, 2026 (npm and GitHub release)
Previous version we covered2.1.316, October 31, 2026
Installnpm install -g @anthropic-ai/claude-code@2.1.323

Check before upgrading

  • Unknown model IDs now assume a 1M context window. It was 200K. If you run Claude Code on another vendor's model with a smaller window, sessions can fail every turn once full until /clear. Set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 before that happens, or use /autocompact to compact earlier.
  • MCP_SDK_GENERATION=v1 no longer gives you the v1 client. The v1 MCP client is removed; that setting now runs the v2 client in a compatibility mode with an earlier handshake and relaxed OAuth checks.
  • ALL_PROXY alone is ignored for organization settings and policy requests. Set HTTPS_PROXY if your network needs a proxy for them.
  • Sandbox writes through symlinks stop working. On macOS, Linux and WSL, write access granted through a symlink that leads elsewhere is ignored; grant the real path. git worktree add, move and remove also fail read-only inside the sandbox.
  • Project MCP approvals match exact names. An .mcp.json approval written in another spelling will ask again.
  • Cross-session messaging limit raised. Sessions on the same machine can send each other about 100,000 messages without a prompt, up from about 20. Teams that relied on the old cap as a brake should know it is effectively gone.

What's new

  • Code Mode (experimental). Claude writes JavaScript that calls its tools instead of calling them one at a time. It is off by default; the Code Mode docs explain how to turn it on, or off for an organization.
  • Plugin version ranges. plugin.json accepts engines["claude-code"]. Plugins outside the range are not installed, updated or loaded, and installs pick the newest version that fits. claude plugin validate also type-checks a plugin's TypeScript hooks, with --strict failing on findings.
  • Usage-credit menu. When a model needs usage credits that are off or used up, you get options to stop, switch to credits or upgrade.
  • Hook heartbeat. Long-running hooks in claude -p and SDK sessions send a progress heartbeat every 30 seconds, so hosts such as Claude Desktop do not end the turn as hung.
  • Smaller additions. .docx uploads for the Artifact tool, a startup warning for Bash allow rules with * inside a variable value such as Bash(PORT=* npm run dev), and CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES to turn off subagent progress summaries.

Permission and sandbox fixes worth knowing

  • --permission-mode auto now starts in Manual mode when auto mode is unavailable, instead of a mode that could allow more.
  • Bash allow rules with an attached wildcard such as Bash(ls*), or combined sudo flags, could skip the prompt in some cases; deny and ask rules could miss commands run through sudo or env. Both are fixed.
  • Write and Edit could write to a path covered by a Read deny rule, and sandbox.excludedCommands could let unlisted commands run outside the sandbox. Both are fixed.
  • claude -p and the Agent SDK ran Bash unsandboxed when sandbox settings failed validation; they now stop at startup and name the bad setting.

Other notable fixes

  • Running /compact twice, or after /context, no longer summarizes away a message Claude had not answered yet.
  • MCP tool calls that fail with the sign-in server's error, such as TOKEN_EXPIRED, now ask you to sign in again.
  • Several cases where a background session moved to a different sign-in or account are fixed; affected sessions now stop and say so.
  • Edit and Bash no longer run out of memory on certain /proc files, devices or named pipes.

Questions

What is Claude Code Code Mode? An experimental, opt-in mode in which Claude writes JavaScript to call its tools. It is off unless you turn it on, and organizations can turn it off.

Why does my non-Anthropic model fail after upgrading? Unknown model IDs now get a 1M-token context assumption. If your model's window is smaller, set CLAUDE_CODE_DISABLE_1M_CONTEXT=1.

Is 2.1.323 the stable npm release? Yes. On November 13, 2026, npm's latest tag pointed to 2.1.323.

More on the Claude Code hub, the previous Claude Code 2.1.316 launch and the coding agents category. For comparisons, see Claude Code vs Codex and Aider vs Claude Code.

Alternatives

Claude Code compared with its alternatives
ToolLicensePricing note
Claude CodeThis launchClaude Code is an agentic coding tool that runs in your terminal and works on your codebase.
—Not stated
Codex CLICodex CLI is OpenAI's Apache-2.0 coding agent that runs locally in your terminal, signed in with a ChatGPT plan or an OpenAI API key.
Apache-2.0The CLI is Apache-2.0. Usage comes from a ChatGPT plan or is billed at API rates when you sign in with an API key.
OpenCodeOpen-source coding agent from the anomalyco/opencode repository, MIT-licensed, with a terminal client and a site at opencode.ai.
MITThe client is MIT-licensed. Model access is separate; this page does not quote a hosted plan.
Gemini CLIGoogle's Apache-2.0 terminal coding agent for Gemini models. Version 0.67.0 (November 4, 2026) adds a models command and sandbox IDE diffs.
Apache-2.0The CLI is free under Apache-2.0. Model use is set by how you sign in: a free Google account tier, a Gemini API key (free or paid), Vertex AI, or a Google AI Pro, Ultra or Code Assist plan.

More from Claude Code

View all

Questions

What is Claude Code?

The repository describes it as an agentic coding tool that lives in your terminal, understands your codebase, and handles routine tasks and git workflows through natural language commands.

Is Claude Code open source?

Not stated in the cited sources. The GitHub API does not report a recognized license for the repository.

Does Claude Code support MCP?

The docs overview lists connecting your tools with MCP, and the v2.1.295 release notes include MCP fixes.

Filed on Agent Radar. Claims in this note are limited to the sources above.