Instant delivery

Lifetime updates

1 domain license

Addonv6.0.2

AI Support Agent for Bicrypto

Name: AI Support Agent for Bicrypto
Brand: MashDiv
SKU: ai-support-agent
Price: 149.00 USD
Availability: InStock

Answers your tickets from your own documentation, cites the passage, and refuses whatever it cannot ground.

2 sales

$149$299Launch −50%

Launch pricing−50% right now

This is a new release, so the first 50 customers get it below its normal price of $299. No code needed — the price you see is the price you pay. When the seats are gone the price goes back up.

10 seats left at this price0 of 50 claimed

10 seats at $14910 left
20 seats at $189−37%
20 seats at $219−27%

You save $150 at today's price. Every licence bought at launch is a full licence — same product, same updates, same support terms as it will have at $299.

You'll also need Bicrypto

On its own, AI Support Agent for Bicrypto won't work — it only runs if you already own Bicrypto. Open the Requirements tab to see exactly what you need and add it to your order.

See it working Live demo Documentation View on CodeCanyon

Answers tickets and live chat from your own articles and your product docs
Shows the passage behind every answer, so a customer checks it and you audit it
Refuses below a retrieval floor instead of guessing, and hands over honestly
Your fees, limits, timings, countries and refunds answerable only from your writing
Offers the answer while the customer is still typing the subject line — no model call
Triages every ticket into a real priority and category, even in copilot mode
Serves your published articles as written, with no answering model call behind them
Optional multilingual answering, grounded in your English documentation
An Assistant that answers YOU about running the platform, and holds actions for approval
Coverage report: what share of your customers' real questions you can answer, and what to write next
Three autonomy rungs, and a measured gate before the console will promote one
Fifteen seeded onboarding questions turn the empty state into a to-do list
Fifteen named handover reasons, sorted into incidents, homework and by-design
Gaps grouped by meaning, so one article closes six phrasings of one question
A nightly job drafts articles from the replies your agents already typed
Hands over with a three-line brief and a suggested reply your agent can edit and send
21 walkthroughs that highlight the control on the customer's own screen
Daily and monthly dollar ceilings that stop it dead rather than overspending
Read-only account tools; passwords, keys, wallets and KYC never leave your server
MashDiv AI metered from your store wallet, or bring your own provider key

Inside AI Support Agent

An assistant you can leave switched on

It answers tickets and live chat from your own knowledge base and the documentation of every product you own, and shows the passage behind every answer. What makes it safe to run unattended is the list of things it will not say — and each one of those becomes a to-do item you can close.

In detail

Your support desk gets a first responder that never sleeps and never guesses. It reads the ticket or live chat, searches your own articles and the documentation of every product you own, and either answers with the passage it used printed underneath, or says so and puts a person on it. It ships switched off and is promoted one rung at a time.

How it works

A turn passes eight gates before a customer sees anything: the channel switches, the spend ceiling, whether a colleague has taken the conversation over, your escalation rules, the retrieval floor, the operator-policy test, the post-filters, and a last claim check under the ticket lock. Retrieval runs before the answering model is called, so a question nothing covers is refused for the price of the search rather than the price of an answer — and produces an honest handover instead of an invention. Answers that do get written are scored against the passages they cite, and withheld when too little sits behind one.

Work it does before it answers anything

Not every job needs a written reply, and the cheap ones start earning on day one. As a customer types a subject line, an article that already solves their problem is offered inline — search only, no model call, and the file-anyway button is never hidden, which is what makes it the one deflection figure that counts tickets which genuinely never happened. Every ticket that is filed gets triaged: one cheap call replaces the customer's own guess at how urgent it is, and their free-text tags, with a priority and a category your queue can sort by. Triage runs even in copilot mode, so the desk is better organised before you have decided to trust a single reply.

Where one of your published articles unmistakably answers the question, it is served as you wrote it, with no answering model call behind it. And with multilingual answering switched on, somebody writing in Spanish, Arabic or Japanese is answered in their own language from your English documentation — with the operator-policy gate holding in every one of them, not just English.

The surface you configure

Screen	What it decides
Overview	Cost, outcomes, why it handed over, what it deflected, and the copilot edit distance the promotion gate is measured on
Live Inbox	The conversations it is handling now — take one over, or hand it back, with a three-line brief and a suggested reply
Assistant	Answers you about running the platform, from the operator documentation and this install's live state
Agents	The name, the voice, the model, and the line that tells a customer it is software
Sources	Your articles, your FAQ and the shipped docs, plus a tester that answers why it refused
Gaps	Questions it could not answer, grouped by meaning, with the article that closes each cluster
Actions	3 reversible account actions and 19 multi-step processes, each off until you allow it — both tiers need the MashDiv gateway; the 21 walkthroughs work on any provider
Provider	MashDiv AI, or your own provider key — your own key lives in the environment file, so it needs shell access
Settings	43 keys: budgets, confidence floors, retention, office hours, which channels it may touch

It answers you, too

The Assistant is the half of this product that never speaks to a customer. Ask it how to run the platform and it answers from the operator documentation and from what this install is actually doing right now — what is waiting in the queue, how the platform is holding up, what your admins have just changed. It can walk you to the control on your own admin screen, and it can hold an action for your approval rather than taking it.

Every catalogue it draws on is filtered by the permissions of whoever is asking, so it can never name a screen, count a queue or offer an action that person could not already reach on their own. It is off until a Super Admin turns it on.

What you can measure before you trust it

The coverage report replays your customers' real question subjects against your own knowledge base and tells you what share of them you can answer today — and which handful of missing pages would move that number most. Every night, the replies your agents typed by hand are turned into draft articles for review, so the questions you answer twice become questions that answer themselves.

What operators control

Autonomy has three rungs — copilot, tickets only, everything — and the console will not promote until it has measured your own agents editing its drafts. Demotion is instant and never questioned. Daily and monthly dollar ceilings stop it dead instead of degrading quietly. Account tools are read-only, and any field carrying a password, an API key, a wallet address or a KYC document is stripped before anything reaches a model. Twelve settings, the master switch and both spend ceilings among them, are Super-Admin only.

AI Support Agent is an addon for Bicrypto and needs no other addon, and a fresh install answers almost nothing — which is correct, not broken. The shipped documentation describes the software, not your fees, limits, review times or which countries you serve — and that is what tickets are about. Fifteen policy questions are seeded blank; filling them is the difference between a demo and a product. Live chat also ships off: tickets only, until you turn it on.

It ships off, and the console will not take your word for it

Autonomy has three rungs: copilot, where a person presses Send on everything; tickets only, where live chat still goes to a human; and everything. Copilot is the shipped setting, and promoting past it is measured rather than asserted — the console counts reviewed drafts, how many were sent, how far your agents rewrote them, and whether your own onboarding questions are answered. Miss any of it and the promotion is refused with the reasons named.

Going back down a rung is instant and never questioned. The master switch, the autonomy mode, both spend ceilings and the account-tools switch are Super-Admin only — twelve keys in total that an ordinary admin can see but not change.

You know what each answer cost, and the ceiling stops it dead

Every turn records what it used and what it cost — tokens in and out, what the prompt cache saved, the price, and how long the customer waited — against whichever provider you run it on: the managed connection or your own provider key. Ceilings ship switched ON at five dollars a day and a hundred a month, on UTC windows, so an install is capped before anybody remembers to cap it. Reaching one does not buy a cheaper model or a shorter answer: it stops, and every conversation escalates to a person.

Set either ceiling to 0 and it is uncapped — the ceiling is your decision, not ours. The one thing it will not do is spend quietly: an install whose provider returns no price for a call is refused outright rather than billed as free.

The cheapest ticket is the one nobody files

As a customer types a subject line, the assistant searches your published articles and offers the one that unmistakably answers it — no model is called, so it costs nothing and is safe to run on a keystroke. Only text you wrote and published can appear there: the shipped documentation cannot, and a draft is not in the index at all. The customer says whether it answered them, and the three outcomes are recorded and counted apart — shown, answered it, filed anyway — because filed-anyway is the one that finds an article which looks right to search and reads wrong to a person. Separately, you can state your team's real hours, and every “someone will reply here” then carries the next opening time in your own timezone.

Nothing runs under twelve characters, and the button that files the ticket anyway is never hidden and never dimmed. The pre-ticket panel ships on; office hours ship off with the schedule already filled in, because hours that are wrong are worse than no hours. The three outcomes are written per offer and returned separately by the overview endpoint over the same thirty days — in this build no admin screen draws them yet.

Every handover says what it means, and which ones you must leave alone

Fifteen named reasons cover every way a conversation reaches a person, and the report bands them before it ranks them: something is broken, this is your homework, this is the guardrail working. A provider outage sits at the top with a Test the connection button. Questions you have not documented come next, broken down by policy topic and matched to the empty seeded article that would close them. Money, account compromise and legal sit at the bottom marked Working as intended and carry no action at all, because there is nothing to fix — five of the fifteen are like that. An outage is rarely the loudest row, so ranking by volume alone would sink the one thing you must act on today below the homework and below the rows you must leave alone.

The window is thirty days on the screen; the API takes one to three hundred and sixty-five, and there is no control for it in this build. A reason nobody mapped still gets a row rather than vanishing — a report that adds up to less than the total it prints is worse than one with an unfamiliar word in it.

Every question it could not answer is a to-do item, not a dead end

A refusal is recorded as a gap, and the gaps are grouped by what customers meant rather than how they typed it, so six phrasings of one question read as one item worth a morning instead of six that each look like noise. Draft an article against the cluster and all six close together. A nightly job does the same thing unprompted, turning replies your agents already typed into drafts for review, and a coverage report replays your customers' real subject lines against your index to tell you which handful of pages moves the number most.

Serving a published article verbatim, with no model call at all, ships off — and in this build there is no control for it on the Settings screen: the key is accepted by the API and by a database row, and nothing renders it. The FAQ & Knowledge Base addon is optional — it adds a third source, it is not required.

It can do things for a customer, and it asks first every time

Three account actions — resend a verification email, withdraw a pending verification, mark notifications read — offered in the conversation and run under the customer's own session when they tap. The assistant holds no credential and no tool it can call accepts a user id, so it cannot act on an account other than the one it is talking to. Multi-step processes show you every step before you approve, because approving a plan and approving a direction of travel are not the same thing.

Actions and processes need the managed MashDiv connection; the 21 walkthroughs work on any provider and are never gated this way. Every one of the three actions is reversible — that is why they are the only three.

Almost nothing about a customer is read, and the words have an expiry date

An account read is deny-by-default: the query names three columns off the wallet and three off the verification record, and the balance is rounded to two significant figures before it goes anywhere. Twenty-two field names — passwords, keys, seeds, addresses, hashes, metadata — are stripped from every result afterwards as a second net, and no tool schema accepts a user id, so there is no account it could read but the one it is talking to. Free text gets different treatment, because prose has no columns to strip: before a conversation is sent to be summarised, harvested into a draft or taught from, twelve patterns run over it in a fixed order — the loosest last, so a narrow rule is never eaten by a broad one — and replace an address, a hash, a card, a phone number or a recovery phrase with a label. Conversation text is then cleared on your own schedule, ninety days by default, and the emptied turn is only removed once the month it was billed in has closed, so the sweep can never hand spend back against your own ceiling.

Switching the account tools off is one Super-Admin toggle, and then nothing the platform knows about a customer account is read into a prompt at all. Retention runs from one day to ten years, and what survives it is the conversation's own record — why it stopped, and what it cost.

gates before an answer is sent: 8
handover reasons, in three bands: 15
walkthroughs it runs on screen: 21
topics only your articles answer: 5

Everything included

134 capabilities, in 11 areas

Every item below exists in the source you receive. Nothing here is a roadmap.

Answering a customer

What happens between a customer writing and a reply appearing.

Support tickets answered, or drafted for a person, per autonomy rung
Live chat in the widget, on its own switch
Every filed ticket triaged into a real priority and category, replacing the customer's own guess and their free-text tags — and it runs in copilot too, so the queue is better sorted before you trust a single reply
Eight gates every turn passes before a customer sees anything
Replies streamed into the thread as they are written
A burst of typing debounced into one answer rather than three
The passage behind each answer quoted in the conversation itself
In-platform links resolved against a live catalogue of your own routes
Thumbs up or down on any answer, with an optional comment
A Talk to a person button that aborts a generation already running
An AI disclosure line carried in the assistant's own voice
A second question sent mid-generation retried, never dropped
Answers in the customer's own language, grounded in English sources

What it will not say

The refusals are the product. Each one is a named, tunable gate.

A retrieval floor below which no model is called at all
A groundedness score checked against the drafted answer before it is sent
Five policy topics answerable only from articles you wrote: fees, minimums and maximums, processing times, countries, refunds
A refund-promise tripwire that discards the answer rather than editing it
A post-filter that strips links the model invented
Built-in triggers that cannot be switched off — money loss, account compromise, legal language, an explicit request for a person, frustration — the first three refusing without even drafting
A turn limit, 1 to 20, after which it hands over regardless
It never names the software vendor to your customers
Customer text and account fields escaped against prompt injection
Escalation rules over seven match types and five actions, three seeded
A glossary of your own terminology carried into every prompt

Knowledge and retrieval

What it is allowed to read, and how a question finds it.

Your own published articles, weighted above everything else
One shipped documentation pack per product you own, indexed with no network access
Your FAQ, when the FAQ & Knowledge Base addon is installed
Scoping to installed and enabled products, with a ten-name alias table
A BM25 keyword index built in-process
Semantic rescue built from the corpus itself — no embedding model, no credential, no extra service
Model re-ranking of retrieved passages, switchable
Passage count per question configurable from 1 to 20
Your live fee, minimum, maximum and processing-time configuration read as citable fact
Current store-hosted documentation fused in by rank, on a 2.5-second timeout
Operator pages excluded from customer answers, and the only corpus an operator question reads
Customer help pages weighted above operator guides
A retrieval tester: what was found, its score, the floor, and whether it would have answered
Per-source chunk count, weight, last-index time and error
Re-index on demand or nightly, with sources replaced atomically
Release notes and resolved tickets never enter the index

Closing the gaps

Every refusal becomes a to-do item, and most of them get drafted for you.

Every unanswerable question recorded with its count, best score and first-seen date
Gaps clustered by meaning, so the count is customers rather than phrasings
Weak passages: what was in front of the model when somebody pressed thumbs-down
A coverage report replaying your own recent ticket subjects, with no model call
Each coverage cluster priced in percentage points of coverage
Inline publish that writes the article and rebuilds the index in seconds
A batch drafter over the top open gaps, re-retrieved against today's corpus
Three named drafting outcomes, including “only you can answer this”
A nightly harvest turning replies your agents already typed into drafts, at no cost
Teach from this ticket: the reusable pattern, never the transcript
Conversations redacted before a model sees them, with the count of identifiers reported back
Only published articles are indexed — a draft cannot reach a customer
A provenance stamp on every machine-written article
Fifteen seeded onboarding questions, shipped blank and undeletable
A closed gap's question promoted to a one-tap starter in the chat
Article create, edit, publish and delete, each rebuilding the index

Handover and the Live Inbox

The screen with a customer waiting on the other end of it.

A full-screen inbox: the queue, the conversation, and the reasoning rail behind each answer
Queue ordered risk first then age, across six session states
Fifteen handover reasons, each with what it means and, where there is one, the next action
Sorted into three bands: incident, your homework, working as designed
A handover report over any window from 1 to 365 days
Policy refusals broken out by topic and matched to the seeded question that closes them
A three-line brief — wants, established, blocking — written once per conversation
A suggested reply loaded into the composer with one press, and never sent for you
Take over, which aborts a generation in flight and marks the conversation yours
Release, or release with the assistant left off for that conversation only
A custody block naming who holds the conversation right now
Automatic reclaim of an abandoned handover, for five reasons only — never when the customer asked for a person
Drafting that continues after a colleague takes over, on a switch
An append-only log of every state change, actor and reason
Notifications narrow by default: the assigned agent, not all 34 people holding the key
Cancel a running multi-step process from the case pane

What it can do for a customer

Three tiers, and the two that can change an account ship off.

Three reversible account actions: resend the verification email, withdraw a pending verification submission, mark notifications read
Constant consent text per action — one tap, or a modal for the consequential one
The model only proposes; the write runs under the customer's own session
Nineteen multi-step processes, with every step shown before you allow one
Twenty-one walkthroughs that open a page and point at the control, one stop at a time
Four of them hand off to a terminal's own onboarding tour rather than duplicating it
Every walkthrough pre-flighted against that customer's own verification level
Two allowlists pointing opposite ways: processes opt-in, walkthroughs opt-out
Nine model tools, three of them account reads
No tool schema accepts a user id — every read is the signed-in customer's own
Docs-only mode: account reads withheld from the prompt, not merely refused
Offers expire after 24 hours and abandoned processes after seven days, swept hourly

Where it appears, and when

The channels, the opening line, and the hours your humans actually work.

Tickets and live chat as two independent switches
A templated greeting with the customer's first name — no model call, no cost, no conversation opened
Up to six one-tap starter questions under it
Office hours as a week of ranges, with lunch breaks and an IANA timezone
Schedule and timezone validated on save, with the unparseable line named back to you
An offline notice carrying the next opening time, so “someone will reply here” gains a when
Pre-ticket deflection: the article that answers a half-typed subject line, with the file-anyway button never hidden
Deflection outcomes recorded — shown, resolved, filed
Suggestions deduplicated inside 30 minutes, and withheld under 12 characters
A published article served verbatim, with no model call at all

Reaching out first

The one capability that contacts a customer who did not contact you.

A ticket opened, with an explanation, for a deposit or withdrawal that failed
A five-minute settling window, and the row re-read immediately before writing
A dry run showing exactly who would be contacted, with every skip named
Three ceilings: per sweep, per customer per day, and per install per day
Deduplicated by transaction tag, so switching it back on never doubles a message
A safe fallback message when the generated one promises a refund
Counted against the same budget, and stopped dead by it

Provider, models and what it spends

Which service answers, and the ceilings that stop it.

Two ways to connect: MashDiv AI, or your own provider key
Credentials read from the environment file; the screen shows a prefix and never a key
A connection test that runs a real generation and separates missing, rejected and timed out
A cost comparison priced against your own usage rather than a brochure's
Switching provider translates all three model settings into the new vocabulary by tier
Separate models for answering, re-ranking and classifying
Reasoning effort across five levels, and it is billed
Answer length from 1,024 to 64,000 tokens
Daily and monthly dollar ceilings on explicit UTC windows; zero means uncapped
A reached ceiling escalates every conversation and notifies the whole desk
Refused and failed turns counted, because the tokens were still billed
A prompt-cache health ratio, with an hourly alarm — the only detection this failure has
A per-turn ledger: tokens, cache reads and writes, cost, latency, retrieval score, groundedness
Six priced models; one with no price refuses to run rather than running unbounded

The console, and who may use it

Nine screens, and the keys that open each of them.

Nine admin screens, grouped by when you use them rather than what they are
An Ask about this customer rail on the customer record, audited like any staff action
An Assistant that answers how to run the platform — from the operator documentation AND this install's live state, every catalogue filtered by the asker's own permissions
Thirteen permission keys of its own, plus three core support keys it reuses
Twelve settings only a Super Admin may change, shown locked with the reason on the row
Forty-six settings keys, five tabs on the screen, written changed-keys-only
A master switch that refuses to enable an install with no working provider
Forty admin routes and eight customer routes, each with its permission declared
A second view of the same tickets — your existing support desk is untouched

Data, retention and what leaves the install

The three questions a compliance review asks.

Seventeen tables, all of them the addon's own
A retention sweep from 1 to 3,650 days that keeps the handover record and deletes the text
Auto-close for tickets nobody replied to, with the explanation posted before the status changes
A twenty-two-field denylist applied to every account read before it can enter a prompt
A free-text redactor for addresses and hashes written into prose
Balances rounded to two significant figures before a model sees them
Anonymised question sharing, off by default and filtered against the vendor's own vocabulary
A cross-tenant answer cache, off, with three conditions before a turn is even shareable
A per-customer deletion runbook, table by table
Eight scheduled jobs, each visible in the platform's own cron console

Licence: Addon licence, activated in the Bicrypto extension manager. Full source.
Requires: Bicrypto core. No other addon is required.
Runs on: Your existing Bicrypto server. No new process, no new service — its eight scheduled jobs run in the platform's cron worker.
Stack: Node routes in the Bicrypto backend, Next admin screens, 17 MySQL or MariaDB tables. The search index is built in-process — no vector database.
AI service: MashDiv AI, or your own provider key set in the environment file.
Admin surface: 9 screens, plus a rail on the customer record
API: 40 admin routes and 8 customer routes
Access control: 13 permission keys of its own, 3 core support keys reused, 12 settings Super-Admin only
Configuration: 46 settings keys, across five tabs
Catalogues: 3 account actions, 19 processes, 21 walkthroughs, 9 model tools
Integrates with: Core support tickets and the live chat widget — one ticket table, so uninstalling leaves the desk byte-identical. Reads the documentation pack of every addon you own, and your FAQ if you run the FAQ & Knowledge Base addon.
You pay for the model: The licence buys the addon, not the tokens. On your own key you are billed by the vendor; on the managed connection you are billed by us.
A fresh install answers almost nothing: The shipped documentation describes the software, not your fees, limits, review times or which countries you serve — and that is what tickets are about. Fifteen policy questions ship seeded and blank; filling them is what turns it into a product.
It ships off, and so does live chat: The master switch is off, autonomy is copilot, and live chat is off. On the shipped defaults it touches tickets only, and drafts rather than sends.
Account actions need the managed connection: Actions and multi-step processes run on the MashDiv provider only. An install on its own provider key gets the 21 walkthroughs and no actions, however the switches are set. Walkthroughs are never gated this way.
The AI key cannot be typed into a screen: The public settings endpoint returns every settings row, so a credential written there would be world-readable. Connecting a provider means editing the environment file and restarting the backend — shell access, not a form.
Four capabilities have no control in this build: Multilingual answering ships off. Verbatim article answers and the shared answer cache are real keys with no field on the settings screen, and escalation rules and the glossary are live tables with no editor — all reachable only through the API or the database.
Five admin buttons spend money off the meter: Ask about this customer, Teach from this ticket, Draft articles, the Handbook and Test connection each call a model. All five are refused once a ceiling is reached, and none of them writes a row the ceiling is computed from.
It never moves money: No tool writes. The three account actions are reversible, non-financial, and execute only when the customer presses the button under their own session.
It answers your platform, not your other channels: Support tickets and the site's chat widget. There is no email, phone, WhatsApp or third-party helpdesk connector.