Meltyfox Guide

What Meltyfox does and who it helps

Meltyfox is a no-code AI chatbot builder for websites. You give the assistant information about your business, choose its appearance and AI models, and embed the chat widget on your site. Visitors can ask questions in their language. When personal help is needed, an operator can join the same conversation.

It suits small businesses, service companies, travel agencies, tour and event organizers, and teams that repeatedly answer questions about their public information. For example, a travel website can explain an event's dates and listed price, then invite a team member for a question that needs judgment.

• Website knowledge: start with written FAQs, text files, PDFs or individual public web pages.
• Multilingual chat: answer visitors in their language, with optional translation during human conversations.
• Human handoff: invite your own operators through email or supported SMS alerts.
• Usage-based funding: manage an organization USD wallet and inspect actual usage.

The assistant is not a booking engine, payment processor for your visitors, or an autonomous employee. Adding a document does not give it access to private systems, live inventory or the ability to complete transactions. Human handoff means inviting a person into text chat, not making a telephone call. Translating chat does not translate the rest of your website.

Quick start with the AI assistant studio

1. Open the Meltyfox homepage and try the sandbox. No payment is needed for this limited trial.
2. In Appearance, enter an assistant name and greeting, then choose an avatar and accent color.
3. In Knowledge, add a short, accurate business introduction and several common questions with their answers. Apply the content and wait until preparation succeeds.
4. Ask realistic questions in the preview: an easy fact, a follow-up, a question in another language and a question whose answer you did not provide.
5. Choose Keep my widget, shown as the save icon in the preview footer. Register or sign in, verify your email if needed, and finish keeping the sandbox.
6. In the registered Studio, confirm the selected organization and widget, configure installation, publish and test on your website.

Keeping a sandbox during signup preserves that widget's edits and knowledge instead of creating an extra default assistant. Continue in the same browser while claiming it. An expired or missing sandbox may need to be recreated. Ordinary signup without a sandbox creates a default widget.

The sandbox is temporary and has reply, file, import and lifetime limits. Its expiry is shown in the preview footer. Pausing to edit does not trigger the installed widget's inactivity timeout, but the sandbox can still expire or reach its usage limits. The refresh icon starts a new sandbox; it is not the same as starting a new chat inside the widget. Keep work you want to retain before resetting.

On a phone, opening Meltyfox's own support chat covers the available browser screen, keeping it separate from your Studio preview. Tap the message field to type; the chat adjusts to the on-screen keyboard. Use Close chat to return to the page without losing your draft or conversation. This does not change your widget's Studio preview or its standard installation layout.

Creating a new sandbox also has daily limits for your network and for the whole site. Private or incognito browsing does not give your network a separate allowance. If a creation limit is reached, the message identifies the limit and shows when you can try again in your local time. These daily creation counters reset at midnight UTC. An existing, unexpired sandbox can still be used within its own limits; a rejected request to start a new sandbox does not replace it.

Operator teams, installation and ongoing knowledge synchronization require a registered organization. Registered accounts can also review chat history in Conversations and receive emailed statistics through Account settings > Usage reports. A successful sandbox test is useful, but does not replace testing the published widget on your actual site.

Accounts, organizations and widgets

Sign in with your email and password. An organization code is an identifier, not a login credential. Use the password-recovery link on the sign-in page if you cannot access your account.

Account

Your administrative ownership group. Account administrators can manage its organizations and administrator access.

Organization

A business or team boundary with its own shared wallet, settings and access grants. Create a separate organization when billing or ownership needs to be separate.

Widget

One assistant with its own appearance, knowledge, installation configuration and operators. Several widgets can share an organization's wallet.

Operator

A person invited to help in that widget's conversations. Adding an operator does not create an administrator login.

Always check the Organization and Widget selectors before editing. Configuring the team for one widget does not configure another. Studio remembers your last accessible selection, while a direct link can select a specific widget.

The account menu leads to Studio, Billing, Account settings and Organization settings. Account administrators manage account-wide invitations; organization administrators have access only to their assigned organization. Use individual invitations instead of sharing passwords. Operator links grant conversation access, not billing or configuration access.

In Account settings, choose whether to receive product news and announcements. For Usage reports, select the organization, choose Off, Daily, Weekly or Monthly, set the report timezone and save preferences. Reports are off until you enable them. Each scheduled email summarizes conversations, handoffs, unanswered handoffs, AI spending and available balance for your review; it does not include a written AI analysis of conversation topics. These report preferences are separate from low-balance alerts and operator invitations.

Use Organization settings to manage the selected organization's logo and, with the appropriate account permissions, organization administrator invitations. A public operator display name or avatar is separate from the administrator profile used when joining through the Inbox.

Customize your website chat widget

Open Widget studio > Appearance. The preview reflects draft changes so you can test before publishing.

• Assistant name: sets the displayed identity and the name the assistant is told to use. It does not add facts about your business.
• Greeting: the opening message shown to a visitor. Prefer a short welcome that describes what this assistant can help with.
• Avatar: choose a supplied portrait or upload your own image. Use an image you have permission to publish. Managed avatar uploads support still PNG, JPEG and WebP images up to 5 MB, with cropping where offered.
• Accent color and theme: choose a swatch, light or dark appearance, and, in a registered account, an editable hexadecimal color such as #FFA7C9. Check text contrast in the preview.
• Vertical position: adjusts the widget's placement. Test against your site's navigation, cookie controls and mobile keyboard.
• Message placeholder: a short hint inside the visitor's input field.
• Quick action buttons: offer common starting questions. These send visitor messages; they are not private system instructions or transaction actions. Additional button editing is available with a registered account.

Studio's interface language and the visitor's chat language are separate. The widget can localize its wording for the visitor; changing the Studio language does not change your intended business facts. Review the greeting and placeholder in each language important to your audience.

Appearance edits save automatically after a short pause. If you temporarily clear the assistant name, greeting, message placeholder or a quick-action label or message, autosave waits until those required fields are complete. Keep my widget and publishing also require complete fields. A successful save clears its earlier save error; unrelated warnings remain visible.

Draft appearance and model changes are not automatically the live website configuration. Wait for saves to finish, test the preview and use Publish in the installation area when ready. Knowledge preparation and handoff settings have their own save flows; do not assume every change waits for that Publish action.

How to install an AI chatbot on your website

1. Select the correct widget in Studio and open Install on your site.
2. Enter Allowed site origins: up to 20 exact website origins, one per line, not comma-separated.
3. Publish the widget and copy the current generated installation snippet.
4. Add the snippet to your website's custom HTML or script area. To make the widget available throughout a site, use its shared layout or site-wide integration area. Your website must allow custom scripts.
5. Open the public website as a visitor and send a test message. Also test a narrow mobile screen.
6. Check the installation status in Studio. The first public widget load starts background site verification.

Example origins:

https://example.com
https://www.example.com
http://localhost:3000

An origin contains the scheme, hostname and optional port, not a page path. Do not enter /about, wildcards or a list separated by commas. Root and www addresses must each be listed if visitors use both. A different protocol or port is a different origin. Adding an origin authorizes embedding; it does not import the website into the assistant's knowledge.

One widget can serve several domains, but each domain can be assigned to only one widget across all Meltyfox accounts and organizations. Its root, www address, other subdomains and ports count as the same domain for assignment; independent sites on supported shared-hosting domains remain separate. Localhost testing is exempt. Saving or publishing an allowed-sites list reserves those domains, even if the widget is later unpublished. Add only domains you manage. If a domain is already assigned, use the account that manages its current widget; contact Meltyfox support if you no longer have access. Creating another account, purchasing funds or receiving a welcome gift does not override an assignment.

To sell a site or move it to separate billing, remove every address for its domain from the old widget and select the save icon, Save allowed sites. Site access changes immediately, but the widget's appearance draft is not published. The domain can then be added to the new widget, which needs its own installation snippet. If removing every site, first unpublish and wait for active conversations to finish, then save the empty list. Unpublishing alone retains the assignment; removing a widget through account controls releases its sites. Do not remove a live site's addresses until you are ready for its old widget to stop serving it.

Use the exact snippet generated for your selected widget. The public widget identifier and verification token are expected to appear there. Never put a private sync key, application credential or provider secret in page HTML.

Verification and local testing

Public HTML verification looks for the installation in your homepage. If a tag manager or client-side rendering hides it from the initial HTML, open For developers: how do I test it? for the supplied meta-tag or DNS fallback and verification status. Use Retry after correcting a failed check; refreshing status alone does not fix the installation.

Localhost and 127.0.0.1 can be allowed at their exact ports for testing. They cannot establish public-site verification or unlock the site activation gift. Account activation credits can already be used while testing locally. Website verification does not read your site into the knowledge base.

Publishing and unpublishing

Publish makes the selected configuration available to new website loads. Unpublish can remain pending while existing conversations finish. Confirm its final status before removing the widget or organization. Publishing, keeping a sandbox, and paying for usage are different actions.

Build a useful chatbot knowledge base

The assistant answers using the information prepared for the selected widget. Start with clear public facts, not a large pile of overlapping documents. Useful topics include what you offer, locations, opening hours, eligibility, dates, prices and currency, inclusions, exclusions, contact routes and customer policies.

Your knowledge, in your language. Provide business information in the language you prefer. Your assistant uses it to answer visitors in their language, without you having to translate the source material. Test important names, prices and business terms in the languages your visitors use.

Select Try knowledge sample for a compact HoneyZebra example with unnumbered sections: business facts, role and boundaries, rules listed in priority order, answer format and a sample question and answer. It fills the title and text and selects the Assistant instructions role without saving a source. Adapt the example to your business, select Save assistant instructions and test the resulting answers. Existing text is replaced only after confirmation. Keep larger business documents separate as Business knowledge.

Start with a required Source title and choose a Source role. Business knowledge accepts a Webpage or PDF/TXT URL, OR Paste text, or a PDF/TXT upload. A custom URL title, such as About page, stays with the source during refreshes. Below the text, the actions are Try knowledge sample, Upload PDF or TXT, then Apply this text as knowledge. Uploads use your title or the filename when the title is blank. Assistant instructions accepts administrator-written text, not a URL or file import.

Choose Assistant instructions for role, boundaries, tone, answer format and approved answer examples. These sources are included in full on every AI answer, even without matching business knowledge. They apply from the next answer without English preparation or a separate preparation charge; their input tokens still contribute to reply costs. Studio limits all assistant-instruction sources together to 12,000 characters per widget, and normal source-count and knowledge-storage limits also apply. Keep instructions short and consistent. The source list marks them Applied on every answer.

An existing manual source titled Administrator rules is still ordinary knowledge until you explicitly change its role. Expand it, select Edit, choose Assistant instructions under Source role, review the text, and select Save assistant instructions. The source ID stays the same. Reverting it to Business knowledge requires preparation again. Website and pushed sources cannot be converted in place; write and review a separate administrator instruction source instead. Imports, titles and document headings never automatically become system instructions.

To correct a manual text or uploaded source, expand it and select Edit. Its original text and title appear in the form; change either and select Apply this text as knowledge. This updates the same source and prepares the changed version, rather than adding another document. Cancel editing leaves the saved source unchanged. Wait until preparation is ready before testing the new information. If someone else updated the source, reopen Edit to load the latest version. Web imports must be refreshed from their URL; pushed sources must be updated at their original integration.

Choose a source

Manual text

Best for a concise introduction, FAQs and stable business facts. Give each source a specific title, apply it and wait for preparation.

TXT or PDF upload

Best for existing text documents. Files are limited to 5 MB, PDFs to 100 pages and extracted content to 200,000 characters, with additional platform limits possible. Use a text-based PDF; scanned images and encrypted PDFs are not supported.

Public page or direct file URL

Import an accessible HTML page, PDF or UTF-8 TXT file. Inspect the extracted content before preparing it. One URL represents one document: importing a homepage does not crawl the entire site.

Pushed integration content

Your website backend supplies selected public records. Update or withdraw these at the publishing system rather than editing managed copies in Studio.

Public-page imports do not execute JavaScript, log in to websites or follow all links. A page that only renders its information after JavaScript runs may import incompletely. Use a readable public document or a server-side push integration instead. Fetching also respects access restrictions and crawler rules.

Preparation and freshness

Applying or importing content authorizes AI preparation within the server's spending cap. Original text is retained and a prepared English version supports retrieval. Wait for a ready state before expecting the new information in answers. Preparation can cost money for a registered widget. Saving a title or seeing a source in a list is not proof that preparation succeeded.

Search and filter sources by Assistant instructions, Manual, Web import or Pushed. Web imports are pulled from a URL, even when their refresh schedule is manual; Manual refresh describes when they update, not where the content came from. Check original content, preparation state, revision and last update. Assistant instructions use the original text directly, without an English-version tab. If business-knowledge preparation fails, read the error, correct the input or funding problem and retry. For a failed URL refresh, the last successfully prepared version stays available.

Knowledge quality checklist

• Keep one authoritative version of each fact. Remove obsolete copies that conflict with fresh data.
• Use descriptive titles and explicit product or event names. Include dates, currencies and timezones where relevant.
• Separate unrelated subjects and remove navigation noise, repeated footers and promotional filler.
• State restrictions and uncertainty. For example, distinguish a listed price from a final quote.
• Test exact facts, paraphrases and follow-up questions after each important update.

Knowledge is reference material, not a secret store or a permissions boundary. Upload only content you are authorized to use and suitable for answering visitors. Do not include passwords, private customer records or payment-card details. Imported pages and visitor messages do not become trusted administrator instructions.

Keep knowledge current with pull and push updates

Choose the simplest update method that matches how your information changes. New website content is not automatically known merely because the chat widget is installed.

Manual updates

Use manual text edits, new file uploads or a URL source's refresh action when changes are infrequent. Check preparation and test the changed fact afterward.

Scheduled URL pulls

Registered URL sources can have manual, daily or weekly synchronization. Open the source's synchronization controls, choose the local time and IANA timezone, select the weekday if needed, and authorize automatic preparation with a per-run USD ceiling. Check synchronization history and the next scheduled run.

Automatic preparation also needs permission in the widget's model policy, sufficient funds and available budgets. The current compact AI quality screen does not expose every policy field. If synchronization reports that automatic preparation is disabled and you cannot change it in Studio, ask the platform administrator to review that policy; use manual refresh meanwhile. A schedule by itself does not authorize spending.

Pause stops automatic refreshes while keeping the last ready knowledge. Removing a source deletes its local knowledge and synchronization settings, not the original external page. A failed fetch, including a missing web page, does not automatically withdraw the last good version. Remove obsolete sources explicitly.

A refresh webhook

A source webhook tells Meltyfox to reread an existing URL. Create its credential in the source controls and call the returned endpoint from your publishing server after a change. It queues work rather than guaranteeing an immediate updated answer. The webhook does not accept replacement content or an arbitrary destination URL. Rotating or revoking its credential invalidates the old one.

Content pushed by your website

Use Knowledge synchronization to create or rotate a server-side sync key. Give it only to the backend integration. That integration can update a stable source when a record is published or edited, and delete its managed source when the record is withdrawn. Rotating the key requires updating the backend.

For a travel website, an event-publication workflow can send the event's public name, dates, program, listed price and link to its Meltyfox widget. The assistant can use it after preparation succeeds. This requires an integration with the website's publishing flow; Meltyfox does not infer the event from a database it cannot access.

URL-refresh webhooks, per-source content push and full application snapshots are different contracts. Use the one appropriate to your integration, monitor failures and avoid creating a new source ID on every edit. See the developer recipes for request examples and withdrawal behavior.

Choose AI models and understand reply estimates

Open Widget studio > AI quality. Defaults are a useful starting point. Compare models using your own questions rather than assuming a more expensive model always improves your assistant.

Answering

Creates the reply from the conversation and relevant prepared knowledge.

Question normalization

Interprets the visitor's question and prepares an English search query for retrieving knowledge.

Translation

Translates replies or widget wording when needed and supports other approved translation work. English answers delivered in English do not need a separate answer-translation call.

Review customer rates per million tokens and model availability in the current catalog. Sandbox choices can be more limited than registered choices. If a selected model is withdrawn, an allowed fallback may be used and shown in Studio. Provider restrictions still apply; a fallback is not permission to use a disallowed provider.

Provider policy controls whether to use only the listed US-vendor or MFC-self-hosted models, or also allow external providers. Vendor classification is not a guarantee that processing stays in a particular country. Relevant questions, conversation history and knowledge may be sent to the selected providers.

Illustrative reply estimate is a comparison aid, not a fixed price or an invoice. Actual work varies with question length, history, retrieved knowledge, response length, language and model. A display of $0.00 (less than $0.005 USD) means a positive sub-half-cent estimate rounded for display, not free usage. Do not multiply the rounded display value to predict a bill.

The platform also enforces per-operation and daily limits. A blocked operation identifies the affected feature, such as sandbox replies, AI work, website imports, new translations or SMS. A daily-limit message shows when you can try again, in your local time; the underlying daily window resets at 00:00 UTC. Other limits can still apply after that reset. A public sandbox limit does not pause every registered widget, and an import limit does not remove previously prepared knowledge.

Sandbox lifetime allowances and empty wallets do not refill at midnight. Keep a sandbox in an account to continue beyond its lifetime allowance, or ask your administrator to review funding or settings. A request can also be too large for the remaining allowance even before the counter reaches 100%. Shorten it or review the AI settings. Failed requests are not automatically resent when capacity returns.

Reduce AI chatbot costs and manage your wallet

Meltyfox uses an organization USD wallet shared by its widgets and registered Studio previews. Top-ups are one-time payments through the configured checkout, not a required subscription or automatic recharge. Promotional funds are used before purchased funds. There is no recurring free allowance.

AI charges use actual reported input, cached-input, separately priced cache-write and output tokens at the applicable customer rates. Work can include question normalization, answering, translation, knowledge preparation, widget wording, handoff summaries and optional dialog translation. A plain human reply is not an AI answer, but an associated translation or SMS notification can have its own cost.

Your wallet charges use Meltyfox customer rates, not the platform's supplier invoices. Sponsored activity still consumes resources and remains metered even when no wallet deduction is made. Provider cost estimates, customer usage value and wallet deductions are different amounts; they should not be added together.

SMS prices are per segment and destination; longer messages can use several segments. Platform-enabled billable image storage and access can add resource charges. Current knowledge and conversation database storage is not a separate customer charge. Review the current Billing ledger and displayed rates instead of relying on a fixed price in this guide.

A practical low-cost setup

1. Start with the default models and a small, accurate FAQ. Test before upgrading models.
2. Try an economical answering model on your real questions. Compare correctness and handoff behavior, not just price.
3. Use concise, current knowledge. Remove duplicates and obsolete material that can increase confusion and processing.
4. Refresh only when information changes. Prefer weekly over daily pulls for stable content, or trigger updates on publication.
5. Leave optional English storage and operator dialog translation off until needed. Known matching-language text is reused; enabling extra language work can increase cost.
6. Use email invitations where sufficient. Enable SMS for the operators and destinations that need it, and set an organization SMS daily budget.
7. Inspect usage after realistic tests. Change one setting at a time so you can understand the result.

Cached widget wording and already prepared matching-language variants can be reused without another translation call. Repeated configuration changes or preparing new source versions can create fresh work. An unchanged URL refresh avoids repeating AI preparation, although fetching and quotas still apply.

Billing and low-balance alerts

Open Billing for the selected organization. Review the balance, transaction history and available top-ups. A checkout success page is not itself proof that funds have arrived; the confirmed payment must be processed. If a payment is pending, check history before buying again.

Set Low-balance alert in USD and enable the email option, then save wallet preferences. A filled value overrides the platform default; a blank value uses that default. Alerts are not an automatic payment and are not repeated for every message below the threshold.

New organizations do not receive low-balance emails while their wallet has never received funds. Email alerts become eligible after a confirmed top-up, an awarded welcome gift or a manual credit. A pending payment or welcome-gift eligibility alone does not activate alerts. Once funded, a wallet can trigger an alert below its threshold, including at zero after its funds have been used. Sponsored wallets do not receive these alerts.

Two separate one-time gifts can help you try registered-account features without topping up. The account activation gift is awarded after email verification creates a new customer account, to its initial organization. It is limited to one per account and email; adding organizations or widgets does not grant it again. The site activation gift is unlocked by verifying an eligible public website. It is limited to one per email and one per site across accounts, independently of the account gift. Its root domain and www address count as the same site. Local testing alone does not unlock the site gift. Current amounts and award status appear beside your wallet on the Billing page. Each offer has a separate platform-admin amount and daily cap, may be paused, and is subject to availability. Changing an amount does not alter previous awards or retroactively credit existing accounts.

Both gifts are promotional funds in the organization wallet, usable across its widgets and allowed sites for AI replies, translations, SMS and stored resources such as operator avatars. Existing feature, safety and usage limits still apply. Credits are not repayable and do not renew automatically. Promotional funds are spent before purchased funds. Ordinary usage cannot borrow or make the wallet negative: new paid operations stop when available funds are insufficient, while already reserved operations settle or refund their reservation. Manual grants and sponsorship remain separate platform-admin decisions; sponsored usage is still metered and limited.

Insufficient funds or a reached budget can pause paid AI work. Failed or invalid AI operations follow the platform's settlement/refund rules; a wallet reversal is not a refund to your payment card. Customer email amounts use two decimal places, while accounting retains smaller USD fractions.

Set up a chatbot with human handoff

1. Choose the right widget and open On-call team.
2. Enable Human handoff and save the handoff settings.
3. Add an operator with an internal name, public display name, email, preferred notification language and enabled notification channels.
4. Mark the operator active. Choose 24/7 availability or working days, times and timezone.
5. Start a new visitor conversation and request help. Open the invitation on another device or browser, open the conversation, join and reply.

Each widget has an independent operator list. Availability means scheduled and contactable, not necessarily online or guaranteed to respond. A conversation captures its eligible operator pool when it starts, so start a new conversation after adding an operator or changing a shift when testing availability.

All available operators invites the eligible team together. Round robin batches invites successive groups until someone joins or the request ends. Choose the batch size and interval for your team; a zero batch size disables new requests. Cooldown limits repeated requests.

While a visitor waits, the assistant can keep answering. When an operator joins, the assistant normally stays quiet. An operator can use Ask the assistant for a reply during human assistance. Return to assistant resolves the human request and resumes normal AI help. Leaving the last operator can return the request to waiting; it is not the same as explicitly returning the conversation to AI.

Configure the unanswered timeout and optional apology. If nobody joins in time, the request times out, its invitation links stop working and the assistant resumes. A visible request or operator-joined notice is part of the visitor conversation.

Email invitations contain a chat reference, time, category, summary and a restricted conversation link. Treat summaries as aids, not exact transcripts. Operators do not need an administrator account to use their invitation. Do not forward private conversation links: they are scoped, expiring access credentials.

The Inbox lets authorized staff review waiting, active, assigned and timed-out requests, join and manage conversations. The Conversations area supports review of widget history. Check delivery status when an invitation is missing; acceptance by a delivery provider is not the same as reaching a person's inbox or phone.

Activate operator SMS alerts and request a country

SMS is an optional operator notification channel. Visitors continue in website chat; operators reply in the conversation room, not by texting their answer to the sending number.

1. In the operator form, enter an international phone number with +, the country code and the full subscriber number. For example, a US number starts with +1. Do not use a domestic-only number or an extension.
2. Enable SMS for the operator and save.
3. In SMS settings, enable organization SMS alerts, set the organization daily budget and check that the destination country is available.
4. Use the operator's SMS activation invitation action to email the private activation link. The phone-shaped action is an activation invitation, not a voice-call button.
5. The operator opens the link, gives their own consent, requests a verification code and enters it. An administrator cannot supply consent for another person.
6. Refresh the operator list and confirm that the phone is verified, then test a new human-assistance request.

Platform delivery, the organization, the destination country, the sender configuration and operator consent must all permit sending. Budgets, segment limits and suppression rules still apply. Saving a phone number or sending an activation email does not activate SMS.

Before creating or emailing an activation invitation, Meltyfox checks SMS delivery setup and whether the phone has opted out. A blocked invitation explains which setting needs attention and whether your organization administrator or an MFC administrator must fix it. If settings change after an invitation was sent, its activation page shows the reason; choose Reload details after the issue is resolved. No SMS is sent by opening or reloading the link. The operator must still give consent and request a verification code.

If the destination is unavailable, use Request country in SMS settings or the operator phone workflow. You can inspect the request status and any review note. A request does not enable delivery or promise a launch date. Email remains an alternative while the country is reviewed.

Operators can use STOP to opt out. Re-enabling a channel checkbox is not a substitute for the recipient's subscription state. Follow the activation page and SMS help if verification, resubscription or delivery is blocked. Never share the activation link or verification code publicly.

Run multilingual chat with translated human replies

Meltyfox's AI flow uses English for retrieval and answering, then delivers the answer in the visitor's language when translation is needed. This lets one prepared knowledge base support multilingual conversations. Proper names, numbers and business terminology still deserve testing.

Human conversations have two independent, optional settings under On-call team > Handoff settings:

• Store English versions for future statistics: prepares English versions of new human-only messages. This is storage for possible future analysis, not a promise of a current English analytics report. Turning it off does not erase English already produced by the normal AI flow.
• Allow dialog translation: lets a joined operator enable translation and localizes invitation summaries into the recipient's preferred language.

Each operator then chooses Use translation in the conversation room. It starts off on each join or rejoin. While enabled, new messages appear in that operator's preferred language and their replies appear in the visitor's language. Other operators make their own choice independently.

For example, a visitor writes in Chinese while an operator works in Ukrainian. After the operator joins and turns on translation, new visitor messages can be shown in Ukrainian and that operator's replies in Chinese. This does not retroactively translate the entire earlier conversation. Use the invitation summary and original history for context.

Original text is retained and Show original remains available. Matching-language text and existing variants are reused. A pending or failed translation shows the original with a status, so an operator should check before assuming the other person has a translated reply. Reading stored history does not itself start a new paid translation.

These options can incur extra AI usage. Turning off one operator's switch stops their requested translation work, not another operator's requests or the independent English-storage setting. A provider call already running may finish.

Read Meltyfox pages in your language

The language selector translates Meltyfox interface text and this guide. Translations are prepared when requested and reused on later visits. When wording changes, only the changed or new text needs a fresh translation; unchanged text stays available. Preparation can finish in parts, with English shown wherever a translation is not ready. If preparation fails, use Retry or continue in English.

Guide translations are automatic and can contain mistakes; the English edition is the original. Links, installation code and technical examples keep their original values. This does not translate your entire website, change your assistant's knowledge, or alter a legal document's accepted version.

Manage assistant behavior and automatic escalation

In Knowledge, choose Assistant instructions for your assistant's role, boundaries, personality, answer format and examples; use Business knowledge for larger factual sources. Keep administrator handoff policy in On-call team > Handoff settings > Assistant handoff instructions (system policy). Assistant-instruction sources can also supply escalation rules within the enabled handoff mode. Ordinary imported pages, pushed event records and visitor messages are not allowed to redefine that policy.

Every AI answer receives published Meltyfox administrator shared guidance first, then the selected widget's assistant instructions. Compatible rules from both apply; shared platform rules take priority when they conflict. Relevant business knowledge supplies reference facts rather than commands. Shared facts about Meltyfox do not describe your business, and unrelated platform details need not appear in replies. Runtime identity, permissions, privacy protections, translation delivery, enabled actions and spending limits remain enforced.

For a playful brand personality, make the condition and expected response explicit in Assistant instructions: When asked about our name, choose one short approved humorous explanation and do not repeat it in the same conversation. Keep fictional origin stories clearly labeled as playful, not verified history. Test the exact question and paraphrases after saving. Instructions are always supplied, but model compliance is not guaranteed; review important responses.

Assistant-initiated handoff

Off (button only)

Keep handoff under the visitor's explicit button control.

Ask the visitor first

The assistant can offer a human when it infers a need, then request one after agreement. A direct request to speak to a person does not need a second confirmation.

Follow administrator rules

Allow administrator policy to request immediate handoff for defined situations, while other inferred needs normally remain offers.

Both automatic handoff and assistant-driven completion start off. Enable them deliberately, save settings and test representative conversations. Human handoff must also be enabled and an eligible operator available. Normal cooldowns, permissions, notification limits and budgets remain enforced.

Example escalation policy

Call an operator immediately, without asking the visitor first, if the visitor becomes aggressive. Offer a human for a refund dispute or a question that requires checking a private booking. Continue answering ordinary questions from the approved knowledge. Do not promise that someone is available or that a refund will be approved.

This is an example to adapt to your business, not a guaranteed aggression detector. Define observable situations and test calm complaints, sarcasm, mixed languages and ambiguous wording. Prefer a confirmation-based policy when a mistaken escalation would be disruptive.

"Without asking first" skips advance confirmation, not transparency: the normal handoff and operator-joined notices remain visible. If no operator is available, the assistant continues without promising contact. Automatic decisions do not run in an unclaimed public sandbox or while an operator is already assisting.

Conversation completion

Choose Off (inactivity timer only), Complete after a clear goodbye, or Ask before completing. A bare thanks, frustration, silence or an unanswered question is not sufficient by itself.

Assistant-driven completion has a 30-second grace period; a new message cancels it. After that completion, a new visitor message can resume the same chat with its history. Manual or inactivity closure is different and requires a new conversation. The configurable inactivity timer applies to installed-widget chats; active unclaimed sandboxes are exempt from that timer, not from sandbox expiry.

Decisions are made with the normal answer operation rather than a separate analysis call, but additional input and output tokens still count as usage. Model decisions are probabilistic, so test and review them instead of relying on them for safety-critical decisions.

Practical recipes for small business websites

A simple customer support chatbot

1. Add a business introduction and FAQs for services, hours, locations and policies.
2. Keep the default models, give the assistant a clear identity and publish to the intended origins.
3. Add one email-enabled operator with realistic working hours.
4. Test an answerable question, an unknown fact, an unavailable operator and a return to AI.
5. Review real questions and improve the knowledge source when an answer was missing.

A chatbot for a travel agency or event organizer

1. Keep general terms and contact details in stable manual knowledge.
2. Keep each event's title, dates, public program, currency, listed price and cancellation conditions together.
3. Connect publishing to a stable pushed source, or refresh a readable public event URL.
4. Withdraw cancelled or unpublished events from knowledge; updating the website alone may not remove a stored snapshot.
5. Escalate requests needing a private booking check. Do not describe displayed availability as live unless the integration actually keeps it current.

A multilingual support team

1. Set each operator's preferred language and verify the languages with a short test.
2. Enable dialog translation at widget level and ask each joined operator to turn on Use translation.
3. Exchange new messages both ways and compare Show original.
4. Check usage before rolling translation out to all operators. Keep optional English storage off unless you need it.

A frequently changing FAQ page

1. Import the exact public page and check its extracted text.
2. Choose a daily or weekly pull, or send a refresh webhook when the page is published.
3. Authorize preparation and set the source's per-run maximum within the widget's policy.
4. Review run history, keep the last-good version in mind and test a changed fact after an update.

Developer recipes for knowledge synchronization

A developer is needed for server-side integrations, not for the normal no-code widget setup. Use the endpoint returned by Studio. Store credentials on the publishing server and redact them from logs. These examples contain placeholders, not working keys.

Update one source using the Studio sync key

PUT /portal-sync/{widgetId}/sources/{sourceId}
Authorization: Bearer YOUR_PRIVATE_SYNC_KEY
Content-Type: application/json

{"title":"Example event","text":"Public dates, program and listed price."}

Use a stable source ID, for example event-123, when editing the same event. Reusing the ID replaces that managed source instead of creating duplicates. Delete it with DELETE on the same path when it should no longer be known. Automatic preparation must be authorized in the widget policy and funded. Confirm successful preparation before treating the content as available.

Trigger an existing URL refresh

POST /knowledge-hooks/{documentId}
Authorization: Bearer YOUR_SOURCE_WEBHOOK_SECRET
Content-Type: application/json

{}

The webhook credential is source-specific and is not the Studio sync key. A successful HTTP 202 response means queued or already-active work, not completed knowledge preparation. Inspect synchronization history. Paused sources reject triggers; rotating credentials requires updating the caller.

Full snapshots for application integrations

The separate application API uses a signed application JWT with knowledge:read and knowledge:write scopes, not the Studio bearer sync key. Read GET /knowledge/sync before building a fresh snapshot, then send:

PUT /knowledge/sync
Authorization: Bearer YOUR_SCOPED_APPLICATION_JWT
Content-Type: application/json

{
  "expectedVersion": 0,
  "documents": [{
    "externalId": "event:123:en",
    "kind": "event",
    "title": "Example event",
    "language": "en",
    "url": "https://example.com/events/example",
    "contentText": "Approved public facts."
  }]
}

A snapshot replaces the integration's synchronized set. Omitted records are withdrawn; an empty list withdraws all snapshot-managed records, not manual knowledge. Keep external IDs stable. On HTTP 409, reread the current version and rebuild from fresh source data instead of attaching a new version to an old payload.

The snapshot contract accepts page or event records, up to 1,000 documents, 50,000 characters per text and 3 MB of serialized records, subject to preparation and platform limits. Source URLs cannot contain credentials, query strings or fragments. This API does not fetch those URLs. Do not expose the application JWT to browsers.

Troubleshooting and frequently asked questions

The widget does not appear on my site

Check the selected widget, its published status, the installed snippet and the exact allowed origin, including www and protocol. Check whether your site permits the loader script and whether a content blocker or content-security policy blocks it. Test outside the administrator session. Installation and verification are related but separate checks.

The assistant cannot prepare or complete an answer

Meltyfox checks prompt size against the selected model's input allowance before requesting an answer. Published shared guidance, widget assistant instructions and the current question are protected. When needed, optional page context, older conversation turns and lower-priority retrieved knowledge are omitted to fit. This does not delete saved knowledge or conversation history. If mandatory content still cannot fit, the request stops with a specific error instead of silently dropping instructions. Keep instructions concise and use Business knowledge for larger factual documents.

Studio shows an advisory warning in Knowledge and AI quality when shared guidance and assistant instructions alone approach the answering model's capacity. The estimate does not include every runtime rule or the next question. Preview errors can show the measured input size, allowance and counting method. Shorten instructions or select an approved model with sufficient capacity; the platform does not automatically raise your limits or switch to a more expensive model to fit the prompt.

Chat errors distinguish oversized instructions or input, provider availability, incomplete answers, preparation failures and translation failures. If your message was saved, the notice says so and your draft remains available. Sending the same unchanged draft again retries that saved message without adding a second visitor message. If the conversation has moved on, you may be asked to send a new message. A successful team request does not clear an unrelated AI error. The notice clears after a successful AI reply, a new conversation or dismissal.

I changed settings but visitors see the old appearance

Wait for the draft save, publish it and reload the website. Confirm that the snippet points to the same widget you edited. The preview can show a draft while the live website still uses the published configuration.

The assistant does not know a new fact

Confirm the correct widget, source contents and successful preparation. Check synchronization history. A website edit alone does not update a manual snapshot. Remove conflicting old sources and ask a fresh question that clearly identifies the topic. Previous replies are not rewritten when knowledge changes.

A PDF or web page fails to import

Check size, page count, text extraction, public access and crawler rules. Scanned or encrypted PDFs, login-only pages and JavaScript-only content need a readable alternative. Inspect the error before retrying, since import quotas still apply.

The contact-team action is missing

Check the widget's handoff enablement, active operators, working hours, invitation mode and usable notification channels. If you just added an operator, start a new conversation to capture the updated availability pool. Configuring a different widget's team does not affect this one.

The operator is saved but there is no SMS

A saved contact is not a verified subscription. Check the activation invitation, recipient consent and phone verification, organization SMS switch, country availability, budgets and delivery history. Use email while a country request is pending. Do not repeatedly resend an uncertain delivery without checking its outcome.

The assistant stays quiet after an operator joins

That is intentional. Use Ask the assistant for a particular reply, or Return to assistant to end the human handoff and let AI resume normally.

A conversation is closed

An installed widget may close after its configured inactivity period or an authorized closure. Use New conversation for those closures. An assistant-resolved conversation can resume on the next message. A sandbox should not close merely because you paused editing; check its expiry and usage limits instead.

The model estimate says $0.00, but funds are decreasing

A small positive amount can round to two decimal places. Check the less-than note, actual token usage and ledger. Knowledge preparation, translations and handoff summaries can also incur usage beyond the visible reply.

The low-balance default seems ignored

A saved organization threshold overrides the default. Clear the field to use the platform default, enable low-balance emails and save. Check the current balance and recent notifications; the alert is not a repeated message for every low-balance request.

Can Meltyfox book, charge or update another system?

Not merely by learning its web pages. The current widget answers questions and supports human text-chat handoff. Knowledge synchronization supplies facts; it is not a transaction tool or a general action connector.

Is the free sandbox a permanent production plan?

No. It is a temporary, limited environment for trying the assistant builder. Keep the widget in a registered account to install it and manage ongoing usage.

Protect access and customer information

• Use individual administrator invitations and the least access each person needs. Operators should use their own restricted conversation links.
• Keep sync keys, application JWTs, activation links and operator links private. Rotate a sync credential if it is exposed and update the integration using it.
• Publish only knowledge and avatars you have permission to share. Do not treat a knowledge upload as a secure vault.
• Explain the use of AI, human support and applicable data processing in your own visitor notices. Review the selected model providers before sending content to them.
• Check original messages when translations affect important details. AI answers, categories, summaries and translations can be wrong.

Public sandbox conversations can be reviewed by authorized Meltyfox platform staff to understand product needs and troubleshoot the service. Do not enter confidential or sensitive personal information into a public trial. Review uses retained messages; it does not join the conversation, generate new translations, extend sandbox lifetime or recover permanently deleted data. Keeping a sandbox in a customer account removes it from the anonymous sandbox review list.

Unpublishing or removing a widget is not the same as erasing all stored data. Widget, organization and account removal use deactivation safeguards and retain records such as conversations and financial history. Unpublish first and let active operations finish before using removal controls. Account and organization operations require appropriate permissions and confirmation.

This guide explains the product; it is not a contract, legal advice or a new acceptance requirement. The published Terms and Conditions and Privacy Policy contain the applicable platform terms and privacy information.

About Meltyfox and contact

Meltyfox Chat is an AI-powered software platform that helps businesses communicate with website visitors, answer questions using their business knowledge, and connect visitors with human support.

Website: meltyfox.com
Support, privacy, and legal inquiries: info.meltyfox.com@gmail.com

Meltyfox Chat is operated by Nirvana Travel Club LLC. When reporting a problem, share the time, widget name and redacted error, not a secret or private access link.

Launch and ongoing maintenance checklist

1. Identity: correct organization and widget, assistant name, greeting, avatar, readable colors and mobile placement.
2. Knowledge: accurate facts, successful preparation, no obsolete duplicates, and tested unknown-answer behavior.
3. Installation: published configuration, exact origins, correct snippet and verified public installation.
4. Languages: test the greeting, a factual answer and a human reply in the languages your customers use.
5. Human support: active operators, realistic schedules, successful invitations, an unanswered timeout and a working return to AI.
6. Automation: deliberately chosen escalation and completion settings, tested with both clear and ambiguous examples.
7. Funding: sufficient wallet balance, low-balance email preferences, reviewed model rates and SMS budgets.
8. Privacy: approved knowledge, protected credentials, individual access and appropriate visitor disclosures.
9. Review: inspect real conversations and usage, fix missing facts, check synchronization failures and repeat the relevant tests after changes.

For ongoing improvement, maintain a short test set of your most common questions, difficult follow-ups, multilingual examples and handoff cases. Re-run it after changing knowledge, models, instructions or integrations. Review costs alongside answer quality and successful human assistance, not message volume alone.