Keys out of shell history and a /status that says what is broken: Claude Code 2.1.327 also closes gaps that let files disappear
The latest version now reads an MCP server's settings from standard input and warns that some commands will ask for permission again.
Anthropic released version 2.1.327 of Claude Code on Tuesday evening, according to the official release page on GitHub. The most notable addition is a new form of the claude mcp add-json command: passing a dash as the input makes Claude Code read an MCP server's JSON configuration from standard input. Access keys therefore no longer have to appear on the command line, where they linger in shell history. According to developer Marc Krenn's changelog archive, this form does not work inside a running Claude Code session.
The /status command also becomes far more talkative. It now lists broken files for agents, skills, commands and output styles, keybinding errors, a sandbox that cannot start, ignored permission settings and hooks whose condition never matches. Claude Code also now asks for confirmation before /install-github-app modifies a repository and before /export overwrites an existing file.
A series of fixes concerns lost data. Release tracker AI/TLDR points out that stopping a Remote Control server in worktree mode could wipe files that git ignores, such as a .env file, from a running session. Project-scope claude mcp add and remove could also discard other keys in .mcp.json, and a self-hosted runner could delete a folder at startup. ExitWorktree now asks first when a worktree may hold unsaved work, and auto mode ignores more allow rules that launch arbitrary packages, such as pnpm dlx. AI/TLDR warns that the stricter path checks mean commands that used to run unprompted may now sometimes ask for permission.
Transparency: this edition was compiled by Claude, made by Anthropic.