New --bare flag for scripted pipelines

Highlight

New --bare flag strips Claude Code down for scripted pipelines

If you call claude -p from scripts, CI jobs, or other automated pipelines, you can now pass --bare to skip hooks, LSP, plugin sync, and skill directory walks entirely. This gives you a minimal, predictable execution path that requires only ANTHROPIC_API_KEY or an apiKeyHelper via --settings -- OAuth and keychain auth are disabled, and auto-memory is turned off. It's a clear signal that Anthropic is treating Claude Code as a composable building block, not just an interactive tool.

Worth Knowing

MCP channels let servers push messages and permission prompts into your session

A new --channels flag (research preview) lets MCP servers push messages directly into your active session rather than waiting for Claude to poll them. Channel servers that declare the permission capability can even forward tool approval prompts to your phone, so you can approve actions while away from your terminal.

VS Code sessions can now bridge to claude.ai for remote control

You can now type /remote-control in the VS Code extension to bridge your active session to claude.ai/code, letting you continue working from a browser or phone. Session titles sync automatically, and /exit properly archives the remote session.

Responses now stream line by line as they are generated

You'll see Claude's output appear line by line in real time instead of waiting for larger chunks. Note that this is currently disabled on Windows (including WSL in Windows Terminal) due to rendering issues.

Rate limit usage now available in your status line

If you use custom status line scripts, a new rate_limits field exposes your 5-hour and 7-day usage windows with used_percentage and resets_at values. Separately, you can now add custom entries to the /model picker via the ANTHROPIC_CUSTOM_MODEL_OPTION environment variable.

Skills, slash commands, and plugin agents now support effort-level frontmatter

You can now set effort, maxTurns, and disallowedTools directly in frontmatter for skills, slash commands, and plugin-shipped agents. This gives you fine-grained control over how hard Claude works and which tools it can reach when a specific command or agent is invoked.

Under the Hood

Startup and resume performance: 60ms faster on macOS, up to 45% faster session resume, ~100 MB less memory in large repos

You'll see faster startup across the board -- macOS gains ~60ms from parallel keychain reads, large repos save ~80 MB on startup, and resuming big sessions is up to 45% faster with 100-150 MB less peak memory.

Session resume, auth, sandbox security, and auto-updater stability fixes

Resumed sessions no longer drop parallel tool results or silently truncate history. Concurrent sessions no longer force repeated re-authentication, compound bash "Always Allow" rules now save correctly per subcommand, and a security fix ensures the sandbox can no longer silently disable itself when dependencies are missing.

From Anthropic

In the News