Claude Code updates — May 22–27, 2026

Claude Code shipped five releases between May 22 and May 27. Here's what actually changed for practitioners — no version numbers unless they help.

---

The five changes worth knowing

1. /simplify is now /code-review (and can auto-apply fixes)

The correctness-checking command was renamed from /simplify to /code-review. The new name is accurate: it reviews the current working tree for logic errors, off-by-one bugs, and inconsistencies — not style simplification.

More importantly: /code-review --fix now auto-applies findings to the working tree. Previously, /simplify surfaced issues but left you to apply them. With --fix, it finds a problem and immediately patches it. This closes the loop on the review-to-fix cycle without a manual step.

If you're using /simplify: update your muscle memory to /code-review. Same command, better semantics.

2. /usage shows you a per-category cost breakdown

The new /usage command outputs a breakdown of where tokens are going in your current session — by category (context, tools, output, cache hits vs. misses).

This is the thing operators and cost-conscious developers have been asking for. Until now, Claude Code sessions were a black box for token attribution. You could see the total at session end, but not whether your cost was coming from context size, tool calls, or output verbosity.

/usage gives you that in-session. Run it mid-session to diagnose why a particular workflow is more expensive than expected.

3. Auto mode no longer requires a consent prompt per action

Claude Code's Auto mode — where it executes actions without asking for approval — previously required a consent confirmation at the start of each session. That prompt is gone.

A background classifier now reviews each action before it executes and flags anything outside the expected pattern for human review. The result: Auto mode sessions run without interruption for routine actions, while genuinely unusual actions still surface.

Practical note: If you're configuring Auto mode for other people on your team, the setup is now simpler. The conversation about what Auto mode does and doesn't require is the same — but the friction of the initial consent prompt is removed.

4. VS Code remote agents survive laptop disconnect

VS Code 1.121 added Remote Agents: agent sessions run over SSH or dev tunnels on a remote machine. If you close your laptop, the agent keeps running on the remote host.

For practitioners running long-horizon agent tasks (multi-file refactors, extended test generation, overnight batch processing), this eliminates the biggest operational risk: a laptop going to sleep mid-task. The session picks up from where it left off when you reconnect.

Setup: Enable remote agent mode in your VS Code settings and point it at a remote host you have SSH access to. The agent runs there; you observe from your local VS Code instance.

5. Enterprise setting: allow all Claude.ai MCP connectors at once

The new allowAllClaudeAiMcps setting in your enterprise configuration unlocks all Claude.ai cloud MCP connectors for your org without requiring admin approval on each individual server.

Previously, every MCP connector needed explicit admin approval before team members could use it. This created a bottleneck at the IT layer. The new setting is an org-wide blanket approval for the official Claude.ai MCP catalog.

Security note: This is an opt-in enterprise setting, not a default. It's appropriate for orgs that have already reviewed the catalog and trust Anthropic's cloud connectors. It should not be used as a shortcut to avoid evaluating individual connectors.

---

One bug fixed that may explain mysterious failures

If your Claude Code setup broke between May 21 and May 23, this is likely why: the May 21 release had a regression where the Bash tool returned exit code 127 on every command regardless of actual result. A hotfix shipped on May 22. If you saw "command not found" errors or unexpected failures in that window, updating to the current version resolves it.

---

What to do right now

Three quick things:

1. Update Claude Code if you're not on the latest version. The Bash exit code regression fix alone is worth it.
2. Run /usage mid-session on your next heavy task. You'll learn something about where your costs are coming from.
3. Try /code-review --fix on a working tree you've been manually patching. The auto-apply is faster than the manual workflow.