Claude Code update lets users keep MCP server access tokens off the command line
Version 2.1.327 of Anthropic’s Claude Code lets users pass an MCP server definition to the program without leaving the access token on the command line. Permission checks also tighten, so unattended jobs should be monitored.
Anthropic released version 2.1.327 of its coding agent Claude Code on 17 November. Most of its more than 170 changes are fixes, the ClaudeKit website reported.
In practice, the most important new feature concerns MCP servers, which Claude Code uses to access external tools. Until now, the claude mcp add-json command has taken the server’s JSON definition from the command line, leaving any access token written into the definition visible. The Claude-news.today website points out that the token can end up in the command history and the process list. If users now type a hyphen after the name, the command reads the JSON from standard input, for example from a file. This does not work inside a Claude Code session, the changelog says. The Claude Media website notes that the official MCP documentation does not yet describe this method.
The update also fixes a bug in which claude mcp add and claude mcp remove deleted other top-level keys from a project’s .mcp.json file, such as another tool’s server definitions. Claude Media advises teams that share the file to check the changes before committing. In addition, /install-github-app now asks for confirmation before it changes a repository.
One change calls for judgement. The first Bash command in a session starts faster because shell start-up files can run before the command is approved. This can happen even when the command is rejected. If start-up files must not run without asking, users can turn the feature off by setting the environment variable CLAUDE_CODE_DISABLE_SHELL_HEAD_START=1.
Permissions are getting stricter. The working directory boundary has changed, so some paths now require permission. Removing a worktree also requires permission if the worktree may contain unsaved work, and some symbolic links are refused outright. If nobody is there to answer the permission prompt, the action is refused. An allow rule, auto mode or bypass permissions mode can still allow worktree removal. Auto mode also ignores more allow rules that permit running any package, and leaves those commands to its safety check. ClaudeKit and Claude Media urge users to check the logs of unattended jobs after updating.
Other fixes stop files from disappearing. For example, Remote Control could delete .env files when the server was stopped in worktree mode, and /export overwrote an existing file without asking.
Read more: How an agent works · Agent map