Support Inbox Agent

support-inbox-agent

A Claude Code skill that does the first pass on your customer support inbox: reads new conversations, classifies them, investigates real issues against your error tracker and codebase, then drafts replies in your email client for a human to review and send.

It never sends anything. Drafts only, always.

Extracted from the production skill we run every morning at 6am on our own support inbox at SiteSpeak.

Read the full write-up: Automate first-pass customer support with Claude Code and MCP

How it works

One markdown file (SKILL.md) describes the workflow. A scheduled claude -p run executes it using four MCP servers:

Role	Example server	Used for
Support inbox	SiteSpeak	list threads, read conversations, pull leads, archive
Error tracker	Sentry	match bug reports to real production errors
App monitor (optional)	Nightwatch	same, second source
Email client	Superhuman	search existing threads, create drafts

Swap any of them for your own stack. The skill is written against tool names, not vendors; the tool-name placeholders are listed at the top of SKILL.md.

Each run:

1. Lists open inbox threads, keeps ones with activity since the last run (a JSON state file makes runs idempotent)
2. Reads the visitor's own messages and classifies the thread: urgent / prospect / noise
3. Investigates before drafting: checks the error tracker and the relevant code for bug reports, verifies how-to answers against your docs
4. Searches your email for an existing thread from that address (people email support minutes after chatting with your bot) and skips or replies in-thread accordingly
5. Creates a draft in your email client
6. Archives the thread and records the outcome in the state file
7. Summarizes: one line per thread, links to drafts

The two rules that matter

From SKILL.md, and worth keeping verbatim:

Never guess. Every claim in a draft must be verified this session or absent from the draft.

NEVER call the send tool. Drafts only, always.

The first keeps confident nonsense out of customer email. The second means the only way a reply reaches a customer is a human pressing send.

Requirements

• Claude Code (or any agent CLI that loads MCP servers and skill files)
• A support tool with an MCP server
• An email client or email API with an MCP server that supports drafts
• Optional: an error tracker MCP server for the investigation step
• A scheduler: launchd (macOS) or cron (Linux). Examples in examples/

Quick start (SiteSpeak)

The example tool names in SKILL.md are the SiteSpeak MCP's real names, so this path needs no adaptation:

1. In Claude Code: /plugin marketplace add sitespeakai/sitespeak-claude-plugin
2. Then: /plugin install sitespeak-chatbot-manager@sitespeak-plugins and authenticate with your API token
3. Copy SKILL.md into ~/.claude/skills/support-inbox-agent/SKILL.md and fill in the placeholders at the top (your chatbot id is in the dashboard URL)
4. Connect your email client's MCP server, and optionally your error tracker's
5. Test manually: claude -p "/support-inbox-agent"
6. Schedule it: see examples/run.sh and examples/launchd.plist

Using another support tool

Any support MCP works if it can list threads, read a conversation, get the visitor's email, and archive a thread. Copy SKILL.md as above, then swap the example tool names for your vendor's equivalents and fill in the same placeholders.

Either way: start with triage only (delete the drafting steps) and read a week of output before you let it draft. Keep sending manual forever.

State

state/processed.json maps visitor ids to { processed_at, action, reason?, draft_id? }. Drafted and skipped threads are both recorded, so reruns never double-process and never double-draft. If a visitor writes again after being processed, the skill updates the existing draft instead of creating a second one.

License

MIT