The plugin does not ship with AI credit. You supply your own Anthropic API key and it is stored on your own site. You pay Anthropic directly for what you use, and nobody marks it up. Nothing AI related runs until a key beginning sk-ant is saved, and without one the desk is still a full helpdesk: it just does not sort, draft or answer by itself.
Getting a key:
- Open console.anthropic.com/settings/keys and sign in, or create a free account. Settings, AI has a How to get your key in 60 seconds drawer with the same link.
- Click Create Key, call it Support Desk, and copy the value that starts with sk-ant.
- Paste it into Settings, AI, in the Your Claude key box, and save. A Connected tag appears beside the box.
What the AI does by itself:
Four switches decide how far the AI goes without you. Start with the first three on and the last one off, read a week of drafts, and switch on auto-send when you trust them.
- Sort every new ticket. Category, mood, urgency and which product it is about, plus how sure it is. The next two need this on.
- Try to resolve it. Licence lookups and how-to answers from your knowledge, only when it is at least as sure as your score. Obvious spam is closed.
- Draft a reply for every ticket. Everything it does not resolve gets a draft. You read it, edit it if you like, and press Send.
- Send drafts without me when confidence is high. Off by default. Licence lookups and how-to or install answers go straight to the customer when the AI is at least as sure as your score. Anything about prices, any question asked before a purchase, and any reply that hands the question to your team still waits for you. No human reads the ones it sends.
How sure must it be to act by itself?
Every ticket gets a score when it is sorted. You set the bar as a percentage: type 90 and it means 90%. At or above it, the AI may answer by itself as the switches allow. Below it, the AI never acts alone and the ticket waits for you, with a draft if drafting is on. The default is 85%. Between 85% and 95% works well, and a higher number means fewer automatic replies. A setting saved before this box became a percentage carries over exactly, so 0.85 shows as 85.
Choosing your models:
Under Which models there are three lists, one per job. Each list holds every Claude model your key can use, newest first, read straight from Anthropic, so a new model is there the day it is released. The newest of each family, Opus, Sonnet, Haiku and Fable, is marked newest, and each option shows its name and its exact model id.
- Everyday replies. Drafts, answers and most replies. A Sonnet is the usual choice. Default claude-sonnet-5.
- Sorting, fast and cheap. Runs on every ticket, so a Haiku keeps it quick and cheap. Default claude-haiku-4-5.
- Hard tickets. Used by itself when a ticket is urgent, angry, a bug, a refund or billing question, five messages or longer, very long, or when sorting was less than 50% sure. Every ticket also has a Draft With Hard Model button to force it. Default claude-opus-5.
Under the lists, a line says when the list was last read from Anthropic. The list is read again every 12 hours on its own, and Check for new models now reads it on the spot. If Anthropic cannot be reached, the desk shows its own built-in list instead and says so. The model you have saved always stays in its list, marked your current choice, even if Anthropic stops listing it, so opening the screen never changes a setting behind your back.
Every model works in every slot, including the newest ones such as Opus 5.5, Sonnet 5.5 and Fable 5.1. Some newer models always think before they answer; the desk knows which, keeps that thinking short for support work, and gives the reply enough room so it is never cut off.
How calls are made:
Requests go to the Anthropic Messages API with prompt caching switched on, so the reusable parts of a prompt, such as your knowledge base excerpts, are sent as cached blocks and cost less the second time. No sampling parameters are sent, because current Claude models reject them.
When Anthropic is busy:
- A moment of overload or a rate limit is waited out and the request is tried again, up to three times, following the wait Anthropic asks for.
- A dropped connection is tried once more.
- A reply that runs out of room is asked for again with twice the space, so customers never get half an answer.
- If a model you chose is retired, the request is sent again on a current model (Haiku 4.5 in place of a retired Haiku, Sonnet 5 in place of anything else), replies keep flowing, and the log says which model was used. Pick a new model from the list when you next visit.
- If Claude declines a request, nothing is sent, the ticket stays with you, and the log gives Claude's own reason.
Every call is logged with the model actually used, token counts, cache read and write tokens, how long it took and any error, in Claude's own words. That log is what the costs below are worked out from.
Your monthly limit:
Spend no more than, in USD a month, defaults to 50. It is a hard stop, not a warning. Before each call the desk works out this month's spend from the log and compares it to your limit. Once it reaches the limit, AI calls stop until next month or until you raise it. Tickets keep arriving normally; only the AI pauses. This month so far is printed under the box.
What the AI costs you:
Below the models, What the AI costs you shows real numbers from your own desk for the dates you pick at its top (Today, Yesterday, 7 days, 30 days, 90 days, This year, Lifetime, or Custom with your own two dates): what was spent in those dates and how many AI jobs ran, what has been spent this month, your monthly ceiling, and what a typical ticket costs to sort and draft. A table lists each kind of job in those dates: how many times it ran, the average cost each and the total. AI costs by day lists every day with AI work, newest first: how many AI jobs ran, what they cost, and a bar against the busiest day. Days are counted in UTC, the clock Anthropic bills by. Under it, a line works out what 100 tickets a month, and 500, would come to at that rate. Answering on your help page is counted separately, because a visitor can ask without ever opening a ticket.
Every figure is worked out per model at Anthropic's own list prices, for input, output and the prompt cache, so mixing a cheap sorting model with a strong hard-ticket model is counted exactly. A model the desk does not have a price for yet is counted at the price of its family, or at the top rate if it is something new, never as free, so the limit can only stop early, never late. Your own bill is on the Anthropic console.