email-me

English · 中文

Monitor, supervise, and steer Claude Code remotely over email. Wherever you are, as long as you can send and receive mail, you can:

• get notified when a task stops, when Claude needs permission/a decision,

or on important events (training finished/failed, monitoring alerts);

• reply to the email with your next instruction and Claude continues

automatically — genuine remote control.

How it works

/email-me:remote on ──arm──▶ Claude stops ──Stop hook──▶ send report email ──▶ poll inbox (blocking)
                                                              │
              you reply from your phone ─────────────────────┘
                                                              ▼
              {"decision":"block","reason": your instruction} ──▶ Claude keeps working ──▶ stops again…

Before you step away, /email-me:remote on to arm (or once to wait just one time). Back at the keyboard, or reply stop / 结束, to disarm and end the loop; a wait with no reply also times out. When not armed (the default), every stop is silent and undisturbed. Arming takes effect from the next stop — the stop that ends the arming command's own turn is skipped, so it never fires an email or burns a once.

Components

| Part | Trigger | Role | |------|---------|------| | hooks Stop | every time Claude stops and armed | sends a report email and blocks waiting for your reply, injecting it so Claude continues | | hooks Notification | Claude needs permission / long idle and armed | sends an alert email (alert only — can't remotely click "allow") | | /email-me:remote | you, manually | arm/disarm: on/once/off, controls whether a stop emails and waits | | skills/notify | Claude judges an event important | proactively emails on training/build done-or-failed, monitoring alerts; can optionally wait for a reply | | /email-me:notify | you, manually | send one alert by hand; --wait to block for a reply | | /email-me:watch | you ask, unprompted | poll the inbox once, treat your mail as a query, gather live state and reply; pair with /loop to keep it running | | scripts/monitor.py | optional background process | continuously persists your replies to inbox/ (fallback for non-blocking mode) |

> Commands carry the plugin namespace prefix email-me:; typing a bare /watch > may report Unknown command when ambiguous.

User-initiated queries (check task progress)

The opposite of "event → notification": with no event triggering it, you email a question at any time, e.g. "How's the training going?". The key point is that answering this needs Claude to look at live state, so inbound queries work by having Claude poll the inbox periodically and reply:

# run the long task in the background, then keep Claude answering (~once a minute)
/loop 60s /email-me:watch

Each tick: scripts/poll_once.py does one non-blocking fetch → any mail from you is handed to Claude → Claude tails logs / reads metrics and answers from real data → scripts/notify.py mails the answer back. If there's no new mail the tick is a no-op.

> Zero-config alternative: run training in the background, let Claude sit at a > Stop, and stop_hook's blocking wait is itself a Q&A window — your question is > injected and Claude's answer becomes the next report email. Good for a one-off > follow-up; for "ask repeatedly, don't end on timeout" use the > /loop 60s /email-me:watch above.

Install

> Requires a Claude Code with plugin support (the claude plugin command). This > repo ships .claude-plugin/marketplace.json, so it can be added as a > marketplace directly.

1. Install the plugin

Project-scoped (enabled only in the current project, good for watching one task temporarily):

claude plugin marketplace add SijiaCui/email-me --scope project
claude plugin install email-me@ohocui-plugins --scope project

Global (available in all projects): drop --scope project (defaults to user scope).

> A local path also works instead of the GitHub repo: > claude plugin marketplace add ./email-me --scope project (after git clone).

After installing, restart Claude Code (or /reload-plugins) so the commands, hooks, and skill take effect; run claude plugin list and confirm

email-me@ohocui-plugins is enabled.

2. Configure environment variables

In the shell that launches Claude:

export EMAIL_BOT=<bot sender mailbox (163), e.g. your-bot@163.com>
export EMAIL_PW=<the bot mailbox's IMAP/SMTP auth code>
export EMAIL_PEER=<your receiving mailbox, e.g. you@example.com>

> Arm switch: the Stop/Notification hooks default to off and are fully > silent — they won't disturb you on every stop. To watch a task remotely, arm > with /email-me:remote on (or once to wait just one time) and the hooks > will email + wait; off disarms. This is the single master switch — no > environment variable needed to enable it. > The SMTP/IMAP servers are hardcoded to 163 (smtp.163.com / > imap.163.com); switching providers means editing scripts/send.py / > scripts/receive.py.

3. Verify

/email-me:notify just testing         # you should receive an alert email

> The first time a script sends mail, Claude Code prompts for permission — pick > "allow". In automated modes it may be blocked; confirm manually, or add rules > to permissions.allow in .claude/settings.local.json: > ``json > "Bash(python3 *)", "Skill(email-me:notify)" > ``

Configuration

| Variable | Default | Meaning | |----------|---------|---------| | EMAIL_BOT | required | the bot mailbox that sends and is monitored (163) | | EMAIL_PW | required | the bot mailbox's auth code | | EMAIL_PEER | required | your receiving address | | EMAIL_WAIT_TIMEOUT | 1500 | seconds the Stop hook waits for a reply (must be ≤ hooks.json's timeout of 1800) |

Limitations

• Notification (permission requests) can only alert, not approve remotely — real

remote control goes through the Stop loop.

• While the Stop hook blocks, that session is frozen waiting; this is expected.
• The bot account is configurable via EMAIL_BOT, but the SMTP/IMAP servers are

hardcoded to 163; switching providers means editing send.py / receive.py.