Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Everyday tasks

The Claude Code email check got narrower in 2.1.321

Written November 13, 2026

2.1.321 narrowed the account-email check to one command shape. What it refuses now, what it no longer stops, and what it never covered.

What changed in 2.1.321

The background is the leak covered in Claude Code sent your email to websites: Claude Code placed the signed-in account's email in the model's context, and the model sometimes copied it into the User-Agent header of requests it wrote. The first response was a privacy mod that stopped commands carrying your email. In version 2.1.321, released November 9, 2026, Anthropic narrowed that refusal. The release notes state the new scope in one sentence: the account email check now refuses only a shell command that holds your email and names a User-Agent, without asking, so other commands and scheduled runs no longer stop at its prompt.

Two halves of that sentence matter. The refusal still exists, and it is a refusal, not a prompt — no approval dialog appears, whatever the permission mode. And the trigger is now the combination: your address plus a word that declares who is calling. A shell command that carries your email for some other reason no longer stops, which is the change that unblocks scheduled runs and scripts that tripped the prompt without ever sending the address to a website.

What counts as naming a User-Agent

The privacy mod's README spells out the matching. The address counts in any letter case, or percent-encoded once. The words that declare a caller include a User-Agent in its spellings (User-Agent, user_agent, --user-agent, UA=), client flags such as curl -A and wget -U, identity words such as set_identity, identity= and mailto=, words for sites that take an address by position, such as the SEC's EDGAR tools, and a From header set as a flag or as a key after the word headers. A computed address counts too: git config user.email in its usual spellings, and the $GIT_AUTHOR_EMAIL and $GIT_COMMITTER_EMAIL variables, when they appear beside one of those words.

To send the address deliberately, the README's route is to set it yourself in an environment variable a script reads, or to run the command yourself. A saved allow rule does not clear the refusal, and neither does the auto-mode classifier.

The mod's wider net, beyond shell commands

The 2.1.321 narrowing is about shell commands. The README describes the same mod watching other surfaces with the same address-plus-word test:

What it never claimed to be

The README's own sentence is the honest summary: this is a courtesy against a well-meaning model, not a boundary. Its list of what passes is long. Your address with no caller word beside it passes — in a request body, in a curl -u argument, in an MCP field such as a calendar invite, or in a file such as package metadata. Values a script reads while it runs pass, because the test sees only the text it is given. An address read on one line and a User-Agent set on another passes. Some tools are outside the test altogether: the address inside a Web search prompt is not covered, and neither are the browser tools or the local read tools. The matching itself is two word searches over the text, with no parser, so it also refuses some harmless calls — the README accepts those false refusals as the price of a simple test.

What to do

Related playbooks

Sources

Back to: Everyday tasks playbooks hub · Also: Claude Code sent your email to websites