Claude Code 2.1.327 shipped November 17 and keeps API tokens off the command line.
A Remote Control bug that deleted git-ignored files like .env in worktrees is now fixed.
Unrecognized models get a 128,000 token reply limit, up from 32,000 before.
About 100 bugs fixed, including /export silently overwriting files without asking first.
On November 17, Anthropic shipped Claude Code 2.1.327, and the two changes worth your attention both stop you from losing something without noticing. One keeps your API tokens out of shell history and process lists. The other closes a bug that could wipe out a .env file you never committed.
Nobody notices a missing file until they need it.
What Changed in Claude Code 2.1.327?
The release note calls it "100 bug fixes," and most of them are small. A few aren't. claude mcp add-json <name> - now reads a server's JSON config from stdin instead of a command-line argument, so a token you're registering never shows up in ps output or your shell history file. /tui classic replaces the old name for the non-fullscreen renderer, and VS Code users get "Claude Code: Add to Chat" in the Explorer's right-click menu, which @-mentions whatever you selected.
The permission system keeps tightening, the same direction it's been moving for weeks. ExitWorktree now asks before removing a worktree that might hold unsaved work, path checks refuse more symbolic links, and permissions.blockReadsOutsideWorkingDirectories catches more Bash commands than it used to.
Auto mode also stopped trusting as many wildcard allow rules that run arbitrary packages, like Bash(pnpm dlx *). If you had a rule like that leaning on auto mode to wave it through, expect a prompt now. It's a narrower door than before, on purpose.
On the performance side, the first Bash command in a session starts faster because shell start-up files can run before Claude Code even approves the command. Windows users get a PowerShell speed bump, somewhere around half a second per command, which sounds small until you've run forty of them in a session. And if you're pointed at a model Claude Code doesn't recognize, a reply that used to get cut off at 32,000 output tokens now continues up to 128,000.
The bug list is where this release earns its keep. claude mcp add and claude mcp remove with project scope used to delete other top-level keys from .mcp.json, things like $schema or a different tool's servers block, if you weren't careful. /export used to overwrite an existing file with no warning at all. Now it asks first.
Neither of those is dramatic by itself. Together with the Remote Control fix below, they read like a team that went looking for every quiet way the tool could destroy a file you didn't ask it to touch.
A handful of the other fixes are worth a mention even if they won't change your day. /rewind used to close without a word when some files couldn't be restored; it now tells you. Pasted images used to get dropped if your prompt started with a file path. Transcript search finally finds text inside edits and written files, not just chat messages. None of those are reasons to update on their own, but they're the kind of papercut that adds up over a few weeks of daily use.
How Does the New Stdin Flag Keep Your Tokens Safer?
Here's the part I'd actually change in my own setup. If you've ever run claude mcp add-json some-server '{"apiKey": "sk-..."}', that key sat in your shell's command history the moment you pressed enter.
On a shared machine, or anywhere a background process can read /proc/<pid>/cmdline while the command runs, that's a real exposure window, not a theoretical one. Terminal multiplexers log it. Some shell-history sync tools back it up to a dotfiles repo without you noticing.
claude mcp add-json <name> - sidesteps all of that. You pipe the JSON in instead: echo '{"apiKey": "sk-..."}' | claude mcp add-json some-server -, or point it at a file. The token never becomes a command-line argument, so it never lands in history or process listings. It's a small flag, but it's the kind of small flag that actually changes a habit.
One limit worth knowing: it doesn't work from inside a running Claude Code session, only from a fresh terminal invocation. That's a minor inconvenience if you were scripting server setup from inside a session, but it's the right tradeoff. I'd rather re-run a setup script once than have a key sitting in scrollback I forgot to clear.
If you maintain any onboarding scripts for an MCP server, whether it's your own or one you're asking others to install, this is worth updating today. It costs one line.
One line, and it closes a hole that's been open since the flag existed.
The same logic applies past MCP setup. Any command you run that takes a secret as a plain argument has the same exposure, whether that's a database CLI, a deploy script, or a one-off curl call you typed by hand. Claude Code just happened to be the tool that gave its own version of the problem a fix this week. The habit worth taking from it is checking where else a key sits in plain text in a place your terminal remembers.
What Was the Remote Control Bug, and Why Does It Matter?
This is the fix I'd have wanted months ago. Stopping the Remote Control server while running in worktree mode used to delete a running session's git-ignored files. .env is the obvious example, and it's exactly the kind of file that's never in git on purpose. If you keep API keys, database URLs or anything else local-only in a .env inside a worktree, stopping Remote Control from your phone or another device could take that file with it, with no warning and no undo.
I run several worktrees at once when I'm testing changes across RAXXO products, and I use Remote Control to check on long sessions from my phone. A typical setup looks like one worktree per feature branch, each with its own .env pointing at a different Shopify development theme or a different local port. I never lost a file to this bug myself, as far as I know.
But "as far as I know" is exactly the problem with a deletion that happens silently. You wouldn't notice until the next time the app wouldn't start because the environment variables it needed were gone.
2.1.327 fixes the deletion itself, and it pairs with the ExitWorktree change mentioned above: the tool now asks before removing a worktree that might hold unsaved work, instead of assuming it's safe to clean up. That's the right layer to add the check at. One fix stops the specific bug, the other catches the next one like it before it ships.
None of that replaces a real backup habit, though. I still keep a copy of every .env outside any worktree, synced to a password manager's secure notes rather than a repo, which is close to the backup habit every RAXXO tool follows before launch.
A tool fixing a bug is good news. Assuming the next bug won't exist is a different thing entirely.
What Should a Solo Builder Do Today?
Update now if you use Remote Control with worktrees. This isn't a release where I'd wait a week to see if anyone hits a regression first. The downside of updating is close to zero, and the upside is a bug that deletes local secrets getting fixed.
Check your version with claude --version before you do anything else. If you're more than a release or two behind, run the update and skim the changelog entries you missed rather than just the one for 2.1.327. Permission behavior has shifted in small ways across 2.1.325 and 2.1.326 too, and jumping straight from an older build can make the new prompts feel sudden when they're really the sum of three releases.
Next, go find any script, alias or README step where you pass an MCP token as a command-line argument, and switch it to the stdin form. I found two in my own setup, both old enough that I'd forgotten they existed.
That's usually how these things get found. Not by auditing on purpose, but by reading a changelog entry and recognizing a bad habit of your own in it.
If you're running Claude Code against a custom provider or a model it doesn't recognize by name, the jump from a 32,000 to a 128,000 token reply ceiling is worth testing. I've hit the old ceiling exactly once, mid-way through a long refactor plan that got cut off before it finished. It's a narrow case, but if you've been quietly working around truncated replies by splitting prompts into smaller chunks, you might not need to anymore.
What I wouldn't do is treat the tightened auto mode permission rules as something to fight. If a wildcard rule you wrote now triggers a prompt it didn't before, that's almost always the sandbox catching something it used to wave through for the wrong reasons. 2.1.325 and 2.1.326 made the same call.
Write a narrower rule. Don't try to get the old behavior back.
Bottom Line
2.1.327 isn't a release built around a headline feature, and that's fine. The stdin flag for MCP tokens and the Remote Control fix both close real gaps, the kind that don't show up in a demo but matter the day they bite you. If you run worktrees with Remote Control, the update isn't optional in any practical sense.
My next step is going through every script I have that touches an MCP server config and converting the token-passing to stdin, then double-checking that my .env backups are current in case the next bug I don't know about yet shows up before Anthropic finds it. Running software that writes to your filesystem on your behalf means living with the small chance it gets something wrong. The honest response isn't to stop trusting the tool. It's to keep a copy of anything you can't afford to lose, update promptly, and read the bug fix section of a changelog as closely as the features section.
This article contains affiliate links. If you sign up through them, I may earn a small commission at no extra cost to you. (Ad)
Top comments (0)