Everything about Granetra

Granetra is an AI chatbot you train on your own content — a website, a PDF, or a Word document — and embed on your site with a single line of code. It answers customer questions in your voice, day and night, using only the material you gave it. When it genuinely doesn't know something, it says so instead of guessing, and it can hand the conversation to a real person.

This documentation covers the full product: creating and training bots, customizing the widget, automation (proactive triggers, agentic actions, conversation flows, AI agent teams, MCP connectors), conversations across the web widget and WhatsApp/Telegram/Slack, analytics, teams, plans and billing, and security. Building against the API instead of the dashboard? Jump to API & webhooks or the full developer reference.

The core loop

1. Train a bot on your content.
2. Customize how it looks and how it talks.
3. Embed one script tag on your site.
4. Monitor conversations and improve it over time.

A bot can be trained from one or more of these at once, and you can add more later:

Pages read per sync is a ceiling, not a quota you spend — every sync starts fresh. A single run also stops at a time budget so training finishes promptly, so a very large site is read as deeply as that budget allows rather than failing outright. Deleting a document in Manage knowledge frees its slot immediately.

Behind the scenes your content is split into passages, turned into embeddings, and stored so the bot can pull the most relevant material into every answer. It never answers from outside what you've given it unless you turn on Flexible reasoning (see Behavior).

Keeping knowledge up to date

• Add knowledge — append more PDFs, Word docs, or URLs to an already-trained bot without starting over.
• Re-sync site — re-crawl the original website URL and replace just the website-sourced knowledge, leaving any uploaded files untouched. Available on every plan. On Scale you can also turn on automatic re-syncing (daily, weekly, or monthly) in Bot Studio → Behavior, so it happens without you.
• Manage knowledge — see every trained source and remove any of them individually from a bot's Knowledge panel.

Everything here lives in Bot Studio → Customize → Appearance:

• Logo — shown on the launcher button and chat header. PNG, JPG, or WebP up to 2 MB.
• Widget color — your accent color, used throughout the panel.
• Color mode — light, dark, or auto (follows your site's own light/dark preference).
• Widget theme Growth plan — 12 ready-made presets matched to common industries (Storefront, Circuit, Care, Estate, Campus, Concierge, Ledger, Counsel, Pulse, Studio, Voyage, Mono). Each ships with both a dark and a light variant, so it follows your site's own theme automatically. Picking one also suggests a matching accent color, which you can change afterwards. Leaving it on Default keeps the standard look.
• Launcher position & bubble style — bottom-left or bottom-right, gradient glow or a flat solid accent.
• Corner roundness, window height & width — sliders, with the panel automatically docking full-screen on mobile regardless of your settings.
• Custom font — match your site's typeface, as long as it's already loaded on the page.
• Custom CSS Scale plan — restyle any part of the widget with your own rules, applied only inside the widget's isolated shadow DOM (it can't leak to your page, and your page can't leak in). Prefix a rule with .dark or .light to target one theme only; Bot Studio includes a live preview with its own dark/light toggle.

• Welcome message — the first thing the bot says when a visitor opens the chat.
• Fallback message — what it says when it can't help.
• Handoff email — where it points visitors when it's stuck.
• Reasoning mode Growth plan — Grounded (default) only ever answers from what you've trained it on. Flexible lets it reason a little beyond your documents for business-related questions, but it still never invents specific prices, dates, or policies, and still declines anything unrelated to your business.
• Visitor memory Growth plan — off by default. When on, the bot remembers returning visitors across sessions, not just within one tab: it recalls durable facts they've shared and a short summary of previous conversations, so a repeat visitor doesn't have to re-explain themselves. Memories are distilled only from what a visitor actually typed, and are deleted along with the bot. Visitors can also erase themselves at any time: when memory is on, the widget shows a quiet "Forget me" link under the message box that wipes everything that bot has stored about them, no account or support ticket needed. It stays visible on white-label widgets too — a privacy control shouldn't disappear because the branding did.
• AI model Scale plan — point a bot at OpenAI (the tuned default), Anthropic Claude, Groq, DeepSeek, Gemini, or xAI. Each provider offers a short, vetted list of models with speed and cost hints, so a typo can't silently send your bot to a model that doesn't exist. Need something not on the list? Use "Request a different or higher model" in the same panel and our team will review it. Embeddings for training always run on OpenAI, regardless of the chat model.
• Live human handoff Scale plan — adds a "Talk to a human" button. See Live human handoff below.

Every bot understands and replies in 95+ languages automatically, detecting whatever the visitor is using — including replying in a language's native script even when a visitor types it phonetically in Latin letters.

Bot Studio → Knowledge lists every source a bot has been trained on — files and pages alike — with the option to remove any single one. To add more, use Add knowledge; to refresh what it knows from a live website, use Re-sync site. Both are covered in detail under Training sources above.

Up to your plan's automation limit of rules per bot (Growth 10, Scale 20), of three types:

• Idle — fires after a set number of seconds on the page.
• Exit intent — fires when a visitor's mouse moves toward closing the tab (desktop only).
• Page — fires after a delay on any URL containing a fragment you choose.

Each rule fires once per visitor per browser session, and only one teaser card shows per page load, so it never feels spammy. Nothing fires while the chat panel is already open, mid-conversation, or during a live handoff.

Register your own HTTPS endpoint as a tool the chat model can call mid-conversation — an order lookup, checking availability, creating a support ticket. The model decides when a tool is relevant, calls it with the parameters you defined, and folds the real result into its answer instead of guessing.

• Up to 6 parameters per action, up to your plan's automation limit per bot.
• GET or POST, with an optional bearer secret sent as an Authorization header.
• A built-in "Test endpoint" button dry-runs a definition against real sample data before you save it.
• Every call is guarded against reaching internal/private addresses, capped at 10 seconds and 4,000 characters of response, and only the parameters you declared are ever sent to your endpoint.

Actions can also be managed over the API — see API & webhooks.

Useful for anything that needs a precise, repeatable sequence rather than a free-form answer — qualifying a lead, booking a demo, routing a refund. A flow starts when a visitor's message contains one of its trigger phrases, and is built from a small set of node types:

Anything a flow doesn't cover falls straight through to the normal AI answer — flows are additive, not a replacement for the knowledge base.

Instead of a single bot answering everything, define up to 8 specialists — Sales, Support, Billing, whatever fits your business — each with its own routing description, optional specialist instructions, its own tool permissions, and optionally its own AI model. Every visitor message is classified and answered by the right specialist, and the widget shows a small label naming who replied.

• Routing is sticky — the conversation stays with one specialist until the topic clearly shifts, so a follow-up question doesn't bounce between agents.
• Tool permissions are per agent — grant a specialist only the agentic actions and MCP connectors it should be able to use. An agent with an empty allowlist gets no tools at all.
• They can hand over mid-answer — if a question clearly belongs to a teammate, the specialist passes it across and that teammate answers straight away, under their own instructions, tools and model. The visitor never has to ask twice and never sees the handover.
• Every reply is attributed — your conversations inbox labels each answer with the specialist who wrote it, including after a mid-answer handover, so you can audit how a conversation moved through your team.
• Specialists add, never override — your knowledge base, safety rules, and grounding still apply to every agent identically.
• Templates — start from a ready-made Sales, Support, or Billing specialist, or create a whole starter team in one click, then edit anything.

If routing can't decide, or you have no agents defined, the bot simply answers as itself — exactly as it does today.

MCP (Model Context Protocol) is the open standard for exposing tools to AI assistants. Connect an MCP server once and your bot discovers every tool it offers automatically — no per-endpoint registration, no schema writing. Those tools are then callable mid-conversation through the same mechanism as agentic actions.

• Up to 3 servers per bot, each contributing up to 25 tools.
• Public HTTPS servers using the Streamable HTTP transport, with an optional bearer token for authenticated servers.
• "Test connection" lists every tool a server exposes before you save, so you know exactly what your bot is gaining.
• Tools are cached when you save, so connecting a server adds no latency to live conversations. Re-saving refreshes the list.
• Each connector can be paused independently, and can be granted to specific agents only when you're running an AI agent team.

Every widget conversation is saved to your Conversations inbox. Filter by bot, or by the "Needs attention" flag, which marks conversations where the bot's answer scored low confidence against your trained content — a strong signal it didn't have a good answer. Open any conversation to read the full transcript, including which replies were flagged.

"Reset analytics" clears a bot's conversation history from your dashboard views without touching your plan's monthly conversation count — usage is measured independently, so it can't be reset by clearing your view.

Connect a channel in Bot Studio → Deploy → Channels and your bot answers there with the exact same knowledge, flows, agent team, and visitor memory as the website widget — and every conversation lands in the same inbox, tagged with the channel it came from.

• WhatsApp — via the Meta Cloud API, using your own Meta app.
• Telegram — via a bot you create with @BotFather. Free and the fastest to set up, so it's the easiest one to try first.
• Slack — answers direct messages and @mentions of your app.

You bring your own credentials, so there's no shared platform account and no Meta business verification to wait on. Each channel has a numbered setup guide in the panel, a "Verify connection" button that checks your credentials live before you rely on them, and a pause switch that stops the channel without disconnecting it.

Live human handoff works on channels too — join a channel conversation from your inbox and your reply is delivered to the visitor in their own app. A returning WhatsApp or Telegram contact is also recognized by visitor memory, the same way a returning website visitor is.

Photos and voice notes are understood, not ignored. Customers on these platforms routinely send a picture of a receipt or record a quick question instead of typing. Your bot reads images — including any text, order numbers or amounts visible in them — and transcribes voice messages, then answers from your knowledge as usual. This works whichever AI model your bot is set to. Files it can't open (PDFs, documents) are acknowledged and the customer is asked for the details, rather than getting no reply at all.

Facebook & Instagram ads: the bot knows which ad they clicked. Point a "Send Message" ad at your connected WhatsApp number and there is nothing else to set up. When someone taps it, your bot opens the conversation already knowing the ad's headline — so it answers about that product or course immediately instead of asking a person who just told you what they want to say it again, which is where most ad-driven chats lose people. It follows them if they change subject, and never mentions the ad or any tracking.

Every one of those conversations is tagged with its ad in your inbox, and Analytics → By ad ranks your campaigns by conversations, engagement, emails captured, and answer rate — how well the bot actually handled the people each ad sent. A campaign with plenty of traffic and a low answer rate usually means the ad promises something your bot hasn't been trained on yet, which is a gap Meta's own reporting can never show you. The ad id and click id are also passed to your CRM with each lead, so you can report conversions back to Meta.

Turn this on in Bot Studio → Behavior and a "Talk to a human" button appears in the widget. When a visitor uses it — or you jump in unprompted — you're emailed instantly, and your replies appear in their chat window within a couple of seconds. Hand the conversation back to the bot whenever you're done, and the visitor can always return to the bot themselves too.

Clicking that button opens a dedicated full-screen chat page in a new tab, carrying the conversation so far across so nothing is repeated. Visitors who are already signed in to Granetra go straight into the chat — they're never asked to log in again. An anonymous visitor is asked to sign up or log in on that page first and then lands right back in the same chat. That takes a couple of seconds for them and gives you a real email address on the conversation — so if they close the tab before you reply, you can still follow up.

Everything stays one conversation: you answer from the same inbox, and the widget on the customer's site keeps showing your replies too, so a visitor can switch between the two tabs freely.

• Conversation volume and answer rate over time, by day or week.
• Top questions visitors actually ask, ranked by frequency.
• Topic breakdown and a simple funnel: started → engaged → resolved.
• Missed opportunities — questions the bot repeatedly couldn't answer, the clearest signal for what to add to its knowledge next.
• Visitor thumbs up/down satisfaction, and leads — moments a visitor was pointed to your handoff email because the bot couldn't help.
• A short, rule-based list of suggested next actions.

At the top of Analytics you'll find a website traffic panel: visitors, pageviews, sessions and pages-per-visit with previous-period comparisons, a timeline you can view daily, weekly, monthly or yearly, your top pages, where traffic came from, device breakdown, and a live count of how many people are on your site right now.

There's nothing to install. The chat widget you've already embedded does the measuring, which also means it measures the pages the widget is installed on — not necessarily every page of your site. If you want full coverage, add the widget site-wide.

It's built to be privacy-respecting by default: no new cookies are set, query strings and URL fragments are stripped before anything is stored (so session tokens or email addresses in your own URLs are never recorded), and referrers are reduced to a bare domain. Traffic data is kept for 400 days.

Invite teammates from Dashboard → Team. Every account starts with its own personal team; inviting someone shares your bots, conversations, and settings with them under a role:

Seats (members plus pending invites) scale with your plan — Growth includes up to 3, Scale up to 10. Every change to a bot, a member, or a setting is recorded in the team's activity log.

All prices shown monthly; paying annually saves 20%. Need something bigger, custom volume, or a signed DPA? See Enterprise.

Every paid plan — Starter, Growth, and Scale alike — carries a verified badge next to your account name in the dashboard, so anyone who sees your profile or admin listing knows it's a real, subscribed Granetra account. Granetra staff (Chief accounts) carry it by default, regardless of plan.

Billing is handled by Polar, our merchant of record — your payment details never touch Granetra's servers directly. Manage your subscription (update your card, switch plans, cancel, view invoices) any time from Dashboard → Manage billing, which opens Polar's secure hosted portal.

• Conversation usage resets on your monthly billing cycle, whether you pay monthly or annually.
• Upgrading takes effect immediately; downgrading takes effect at your next renewal.
• Enterprise billing is invoice-based with a dedicated DPA — see Enterprise.

• Your data stays yours — your knowledge base is only ever used to answer your own visitors. It's never sold or shared.
• Encrypted end to end — everything is encrypted in transit (TLS) and at rest.
• Never used to train AI models — your content and conversations are never fed back into any model's training.
• GDPR-ready by design — export or permanently delete your account and all its data at any time from Settings, no support ticket required.
• Account protection — two-factor authentication and a full active-sessions view (see and sign out any device) are available to every account, free.

Full detail lives in our Privacy Policy and Terms of Service.

A REST API and signed webhooks for automating everything the dashboard does: list your bots, read conversations and transcripts, send a chat message and get a reply, and manage agentic actions — all with a scoped bearer key. Webhooks deliver a signed, retried POST the moment a conversation starts, a visitor asks for a human, or a lead is captured. Keys can also carry an IP allowlist and report per-key usage.

Read the full API reference

The widget doesn't appear on my site

Confirm the script tag is present and placed before </body>, and that the bot's status is "Ready" (not "Training" or "Failed"). If you've set an "Allowed websites" list in Bot Studio → Access, make sure the domain you're testing on is included — otherwise the widget silently refuses to load there by design.

Training fails, or a website won't crawl

Some sites block automated crawlers, and pages rendered entirely with JavaScript can't be read this way. Try a PDF or Word export of the same content instead — it almost always resolves it.

The bot keeps saying it doesn't know

Check that the content you expect it to know is actually listed in Knowledge base. Add more source material with "Add knowledge", or turn on "Flexible reasoning" (Growth+) so it can reason a little beyond your exact documents.

A visitor is asked to sign up partway through a chat

As a usage safeguard, an anonymous visitor who reaches 5 messages in one conversation without leaving contact info is invited to sign in or create a free account to continue. It doesn't apply if the person is a signed-in Granetra user themselves — this only affects genuinely anonymous website visitors.

I've hit my monthly conversation limit

Your usage meter on the dashboard shows exactly where you stand and when it resets. Upgrade your plan at any time for a higher limit — see Plans.

Logged in? Use the "Found a bug?" button in your dashboard sidebar to report a problem directly to our team. For anything else, email granetra@gmail.com — or just open the chat bubble in the corner of this page. It's a real Granetra bot, trained on this very documentation.

Already using Granetra? We'd genuinely like to hear how it's going — copy your Bot ID from any bot card and leave a review on our reviews page.