# intoCHAT
> intoCHAT is a no-code tool for building an AI chat agent trained on your own website, documents and Q&A, and putting it on your website with one script tag. It answers visitors' questions from your content, collects leads, searches live products in online shops and can call your own APIs. The same agent can also answer in Slack and by email, and conversations that need a person land in a shared team inbox.
Website: https://www.intochat.ai (brand name is written "intoCHAT"). Category: SaaS, AI chatbot builder, website chat agent. Fact sheet for AI assistants: https://www.intochat.ai/en/llm-info
## What intoCHAT does
- Training: crawls a website (sitemap and link discovery, up to about 500 pages per scan, a page picker grouped by folder with folder checkboxes, exclude rules, JavaScript-heavy pages rendered), uploads files (PDF, Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, TXT, Markdown), text snippets and Q&A pairs (up to five other phrasings per question; export to CSV; import from CSV, XLSX, XLSM or ODS, up to 1,000 rows per file). Where Notion is set up, it also imports Notion pages through a read-only connection the owner authorizes in Notion (up to 25 pages per import; text blocks are read, images and files give only their caption, subpages and databases are skipped; Notion addresses are never shown to visitors or the model). Where Google Drive is set up, the owner can connect a Google account and pick files in the Google Picker (Google Docs, Sheets with up to 1,000 rows per sheet, Slides, PDF, Word .docx, text and Markdown; up to 10 files per import, 50 MB per file); intoCHAT uses only the drive.file permission, so it can open just the picked files, and Drive addresses are never shown to visitors or the model; on disconnect the owner keeps or deletes the imported files. Agents can also learn from solved Zendesk and Freshdesk tickets through the agent's helpdesk connection (Starter 200, Pro 1,000, Enterprise 5,000 tickets per agent; not on Free): solved or closed tickets from the last 30, 90 or 365 days become question-and-answer sources (the customer's first message and the team's public replies; internal notes and tickets without a public reply are skipped), with emails, phone numbers, card, order and account numbers, token links and names in greetings and sign-offs removed before saving, and long conversations optionally condensed by the AI after that removal; and from published Intercom Help Center articles (US, EU or AU region, optionally one collection) with an access token the owner pastes, stored encrypted. Both can sync daily or weekly. Q&A answers and starred sources rank first. Training runs on intoCHAT's servers and continues after the owner closes the page; interrupted runs are resumed. Optional auto-resync (off by default) re-reads website and Notion pages and Google Drive files (only those modified in Drive) daily or weekly and retrains only those whose text changed; uploaded files, text and Q&A are never touched and nothing is deleted. Each run that reads website pages uses one website crawl and covers up to 30 pages; Notion pages and Drive files use no crawls. Website pages can also be refreshed one by one or per site or folder. The source list can be grouped by site and folder, and sources can be exported as a ZIP of plain-text files with a manifest.csv. Crawl settings per agent: the scraping country (for the crawling provider), a slow mode for fragile sites, and whether pages that differ only by their ?query count as separate pages (tracking parameters are always dropped; query pages are capped). A new agent can also be started from a website address: intoCHAT reads the homepage and up to four key pages, takes the name from the site title and the logo and colours from the site, picks an instruction template, trains the agent and opens it in the playground (uses one website crawl).
- Knowledge conflicts (Pro and Enterprise): after training, once a day and on request, an AI model compares passages from different sources that cover the same thing and lists the facts they disagree on (prices, hours, policies, numbers, dates, names) as contradictions, outdated facts or inconsistent duplicates, with both facts side by side and a suggestion; findings resolve themselves when a passage changes or a source is deleted, and a dismissed pair is not flagged again unless its text changes. Up to 60 pairs per check and 240 per agent a day; checks use no plan messages. Docs: https://www.intochat.ai/en/docs/knowledge-conflicts
- Answers: hybrid semantic and keyword search over your content. The agent is instructed to stay grounded in your sources, to say when it does not know, and to ignore prompt-injection attempts. It replies in the visitor's language unless the owner sets a reply language.
- Setup: instructions (system prompt) with starter templates, guardrails (must-follow rules of up to 4,000 characters, added after the instructions; intoCHAT's own rules still apply), a Stay on topic switch (off by default; the agent politely declines requests unrelated to the business, while greetings and small talk still get a short reply), a reply language (automatic by default, so the agent answers in the visitor's language; or one of 31 languages, used when the visitor's language is unclear or, with "Always use this language", for every reply; separate from the widget language), a welcome message plus up to three more, suggested questions, and a playground that tests unsaved instructions, guardrails and reply language. Hand-off by email (off by default): when a visitor asks for a person, the agent can email the conversation, a short summary and the visitor's address to the lead-alert recipients, who reply by email; at most once per conversation every 10 minutes and 20 times per agent per hour. Optional helpdesk tickets: with Zendesk (API token), Freshdesk (API key), HubSpot (private app token), Gorgias (API key) or Help Scout (app ID and secret) connected, each hand-off also opens one ticket there with the visitor's email as the requester, an AI summary, the last 30 messages and a link to the conversation, plus optional ticket fields (custom fields, tags, priority, group and similar, filled with fixed text or conversation, contact and lead variables); one ticket per conversation, no sync back (replies happen in the helpdesk), and the hand-off email is still sent. Live chat (off by default): when a visitor asks for a person while someone from the team has the Live chat inbox open, a team member can take over the chat in the widget and reply under their first name while the AI stays silent (polled every 3 to 5 seconds, not instant; a request nobody takes within 5 minutes lapses and the email hand-off is offered if on; optional 1-to-5 visitor rating; messages during a takeover do not use the plan's messages). An agent can be duplicated with its settings, knowledge, actions, chat forms and procedures, without retraining; conversations, leads, form submissions, webhooks, preview links and settings history are not copied. Settings history: every saved change to an agent's settings (instructions, guardrails, appearance, lead, access and other settings; not knowledge, actions, forms or procedures, and never secrets) is kept as a version with who saved it and which settings changed, the last 100 per agent; any version can be compared with the current settings (long text line by line) and restored in one step by the owner or an admin, and the restore is itself a version that can be undone. Optional web search (off by default): when the owner's knowledge does not answer an in-scope question, the agent can run one web search per reply through OpenAI, optionally limited to up to 20 sites; the reply says the information comes from the web and links up to 3 cited pages on a "From the web" line, and each reply still counts as one message. Each agent answers with Claude Haiku 5.5 (Anthropic) by default, on every plan including Free; GPT-6 Luna (OpenAI) can be chosen on every plan too, both at one message per reply. On paid plans the owner can choose another model per agent from OpenAI, Anthropic or Google where intoCHAT has that provider set up (Gemini Flash from Starter at one message per reply; Claude Sonnet 5.5 and Gemini Pro at two and Claude Opus 5.5 at five messages per reply from Pro), or Auto, which picks a model per message for one message; the visitor's messages go to the chosen model's provider, and if that provider fails GPT-6 Luna answers. An optional per-agent temperature (off by default, 0 to 1) is used only by Gemini Flash and Gemini Pro (with Auto, only for replies by Gemini Flash); Claude Haiku 5.5, GPT-6 Luna, Claude Sonnet 5.5 and Claude Opus 5.5 set their own.
- Channels beyond the website (Starter and up): the Slack agent answers direct messages to the intoCHAT app and @mentions in public and private channels it was invited to, replying in the thread (one thread is one conversation) with the same knowledge, instructions, actions and rules as the widget, up to 300 messages per workspace per hour; the email agent gives an agent its own email address, which customers write to directly or which the owner's support address forwards to (Gmail, Google Workspace, Microsoft 365, Outlook), and answers each email in its thread; on Pro and Enterprise replies can come from the owner's own verified domain. Email attachments are not read. Slack and email conversations appear on the Conversations tab with a badge, and the team can pause the AI and reply from intoCHAT. Each AI reply counts on the plan as in the widget. Docs: https://www.intochat.ai/en/docs/slack-agent and https://www.intochat.ai/en/docs/email-agent
- WhatsApp, Facebook Messenger and Instagram Direct (Starter and up: 3 channels per agent on Starter, 6 on Pro, 20 on Enterprise; a WhatsApp Business number, a Facebook Page or an Instagram professional account each count as one): connect from the agent's Share tab (WhatsApp through Meta's Embedded Signup or a system user token, Messenger and Instagram through Facebook Login or a Page token), and the agent answers messages people send there with the same knowledge, instructions and actions as the widget. It replies with text and reads text, photos and PDFs (with visitor attachments on); forms, the lead form, booking times and follow-up buttons aren't shown in messaging apps, and action buttons are sent as links. Conversations appear on the Conversations tab and in the Inbox with a channel badge; the team can pause the AI and reply from intoCHAT, and the AI resumes after 24 hours. Following Meta's 24-hour rule, the agent only answers messages people sent: it never writes first and sends no template, broadcast or marketing messages. Each AI reply counts on the plan as in the widget. Docs: https://www.intochat.ai/en/docs/whatsapp-and-meta
- Copilot (Starter and up; 30, 150 or 500 messages a day per account on Starter, Pro and Enterprise): a chat in the agent's dashboard that changes the agent on request (instructions, welcome message, suggested questions, Q&A pairs, answers to unanswered questions, buttons and more); every change comes as a proposal with a diff that nothing applies until a person approves it. Docs: https://www.intochat.ai/en/docs/copilot
- Rich replies (on by default, switchable per agent): the agent can add a table, a row of cards or quick-reply buttons under its text reply, filled only from the knowledge, the conversation or action results, at no extra message cost.
- Widget: one script tag, an inline iframe, a shareable link, a hosted Help Page, the official WordPress plugin and a Shopify theme snippet. Colours, a light, dark or automatic theme, font, corner style, avatar, launcher icon and text, placeholder, footer text and position are configurable; brand import pulls your logo and colours from your website. The widget's own text (buttons, notices, error messages, default welcome and placeholder, lead form) appears in English, German, French, Italian or Spanish: automatically from the page's language, then the visitor's browser, or fixed by the owner; text the owner writes is shown as written. Suggested questions can be a simple list or up to four titled groups shown as tabs. Optional AI follow-up questions: 2 or 3 short questions as buttons under the latest reply, in the reply's language, generated with the reply (no extra message). Optional source links: up to 3 links under a reply to the website pages it drew on, taken from the knowledge the reply was given, never written by the model; files, text snippets and Q&A pairs have no links. The chat bubble can show the welcome messages as a pop-up above the launcher and open the chat by itself after a delay (auto-open runs at most once per visit and not on screens up to 768 pixels wide). Optional previous chats (off by default): visitors can reopen up to 10 of their earlier chats from the last 90 days and continue them; the list is kept only in the visitor's browser (per website for the chat bubble), never in an account. Optional visitor attachments (off by default, available once intoCHAT's private file store is set up): visitors can send images (PNG, JPEG, WebP, GIF) and PDFs, up to 4 MB each and 3 per message, which the agent reads with the message (sent to the provider of the agent's model; PDFs are read only by OpenAI models); files are stored privately with the conversation, shown in the owner's transcript and deleted with it; not in temporary chats. "Powered by intoCHAT" can be removed on paid plans.
- Share: a Help Page per agent (off by default) at https://www.intochat.ai/help/ with a title, a description, up to 6 links and the chat, using the agent's look and widget language; it is not available for private agents, can be indexed by search engines at its intoCHAT address, and can be linked from or framed on the owner's site. Custom domain (Pro and Enterprise, one per agent): the owner points a subdomain such as chat.example.com at intoCHAT with a CNAME record (plus a TXT record when the hosting provider asks for one), and that domain then serves the agent's Help Page at its root (or the full-page chat when the Help Page is off), the chat at /chat and the widget script at /embed.js, whose chat window also loads from the custom domain; HTTPS is automatic, root domains such as example.com are not supported, only that agent is reachable on the domain, and the intoCHAT addresses keep working. Docs: https://www.intochat.ai/en/docs/custom-domain Preview links: a private link (https://www.intochat.ai/p/...) that shows the real agent over a screenshot of a prospect's website, for sales; each new link uses one website crawl, up to 20 open links per agent, optional expiry after 7, 30 or 90 days, view counts that skip bots and link previews, not indexed; chats on it are real and count toward the plan.
- Access: allowed domains limit the websites an agent can appear on (up to 20 domains, subdomains included; an empty list allows any site). A private agent answers only its owner, in the playground. Per agent, the owner can also limit messages per visitor (counted by IP address, 1 to 1,000 per hour, in a fixed one-hour window from the visitor's first message), cap the agent's replies per monthly usage period (the account's plan limit still applies to all agents together), and block the chat in up to 50 countries, judged by the visitor's IP address (a VPN gets around it; an unknown country is allowed). Refused messages do not use the owner's messages. The playground is exempt from the visitor limit and country blocking, but counts toward the monthly cap. Identity verification: the owner's server signs a signed-in visitor's user id (HMAC-SHA256 with a per-agent secret) and the page passes it to the chat bubble's JavaScript SDK (window.IntoChat, which can also open and close the chat, switch the widget language and report chat events); verified conversations show the user id, verified visitors become contacts, server-side API actions can use {{user.*}} placeholders filled from the verified identity, the AI is told only that the visitor is verified and their name, and the owner can let only verified visitors chat.
- Leads: an in-chat form for email and/or phone, plus an optional name and up to five custom questions (short text, long text or dropdown), each field off, optional or required; shown manually, by the AI, by rules (message count, keywords) or both; contact details typed in the chat are also captured; email alerts to up to six recipients; Slack alerts that post new leads, hand-off requests, live chat requests, form submissions and/or booked meetings to one Slack channel through a Slack incoming webhook the owner creates (incoming-webhook URLs only, not Workflow Builder links; intoCHAT only posts to the channel and reads nothing from Slack); CSV export.
- Team inbox (every plan): every conversation that needs a person (hand-offs, live chat requests, spam pauses, actions waiting for approval, or moved there by hand or the REST API), across all agents, with statuses (open, pending, snoozed, solved), priorities, tags, assignment to a teammate, bulk changes and saved views (shared views from Starter). Inbox teams and routing (Pro: 5 teams and 20 rules; Enterprise: unlimited): assignment rules send a conversation to a team, shifts and holidays with an outside-hours auto-reply, and least-busy or round-robin assignment among teammates who are online, on shift and under their limit. Inbox automation: saved replies with variables (Starter and up), triggers that tag, prioritise, set the status, send a message or notify by email or Slack (Pro and up), and live translation so each teammate reads and replies in their own language (Pro and up). Satisfaction emails (Starter and up, off by default) ask the visitor for a 1-to-5 rating after a conversation is solved. Docs: https://www.intochat.ai/en/docs/inbox and https://www.intochat.ai/en/docs/inbox-teams-and-routing
- Contacts: one record per person per agent (Free 100, Starter 5,000, Pro 25,000, Enterprise 100,000 per agent) with email, phone, name, the person's user id on the owner's site and up to 50 custom attributes. Contacts are created from identity-verified visitors (matched by user id only), from new leads (which never change a contact from another source), from a CSV import (up to 10,000 rows and 5 MB, previewed first; other columns become attributes) and from the REST API or MCP server. Searchable list with a detail view showing attributes and conversations, attribute editing, delete and a formula-safe CSV export. Server-side API actions can read {{contact.email}}, {{contact.name}}, {{contact.phone}}, {{contact.external_id}} and {{contact.}}: for a verified visitor their own contact, otherwise the contact of the lead saved in the conversation (never one with a user id); if a value is missing the action does not run. The AI never sees attributes unless an action returns them. Not copied when an agent is duplicated. Docs: https://www.intochat.ai/en/docs/contacts
- Webhooks: signed JSON POSTs (HMAC-SHA256) for new leads (lead.created), hand-off requests (handoff.requested), chat form submissions (form.submitted), booked meetings (booking.created), live chat requests (live_chat.requested), return requests (return.requested), orders by message (order_request.created), inbox status changes and assignments (conversation.status_changed, conversation.assigned) and action approvals (action_approval.requested, action_approval.decided) to up to five https endpoints per agent, with retries and a delivery log. They work with your own server or tools that accept incoming webhooks, such as Zapier, Make or n8n. The Zapier app subscribes these webhooks through the REST API; Make and n8n use their own incoming-webhook steps.
- REST API (Starter and up): with an account API key (scoped, shown once, 60 requests a minute) a server can chat with an agent, each reply counting as one plan message, read conversations and leads, change a conversation's inbox status, priority, tags and assignee, read, create, update and delete contacts with their attributes, add or delete text and Q&A sources, subscribe and remove webhooks for an agent's events (REST hooks, as Zapier and Make use them; revoking the key removes them), and list recent events in the webhook format for polling; it has no endpoints to create or change agents. Docs: https://www.intochat.ai/en/docs/rest-api. OpenAPI 3.1 spec (public, importable into Postman or Insomnia): https://www.intochat.ai/api/v1/openapi.json
- Zapier app (Starter and up, connects with an API key): listed in Zapier's app directory at https://zapier.com/apps/intochat/integrations. 7 triggers (New Lead, New Form Submission, New Booking, New Return Request, Hand-Off Requested, Live Chat Requested, New Conversation), 3 actions (Send Message to Agent, Add Q&A Pair, Add Text Snippet) and 1 search (Find Conversation). Docs: https://www.intochat.ai/en/docs/zapier-and-make. There is no Make app yet.
- MCP server (Starter and up): a remote MCP server at https://www.intochat.ai/api/mcp lets AI tools list agents, chat with an agent (each reply counting as one plan message), read conversations, leads and contacts, create or update contacts, and add or delete text and Q&A sources. Claude.ai and ChatGPT connectors connect by signing in (OAuth with dynamic client registration and PKCE; the person chooses read only or read and write, the account and the agents, never beyond their own role, and can revoke the app under Settings → Account → Connected apps); tools that send a header (such as Claude Code, Cursor or the MCP Inspector) use the same scoped API keys and rate limit as the REST API. Docs: https://www.intochat.ai/en/docs/mcp-server
- Actions: custom HTTP API actions (any method, bearer/basic/API-key auth, cURL import, field-level data access, up to 25 per agent), actions that run JavaScript in the visitor's browser, custom buttons: a link with your text that the agent shows under a reply when it fits (up to 3 per reply; intoCHAT never calls the link), and chat forms: up to 10 owner-built forms per agent (up to 15 fields: short or long text, email, phone, number, date, dropdown) that the agent shows under a reply when they fit, built from the stored form, not in temporary chats, one submission per form per conversation; submissions are listed on the Leads tab with CSV export, are not leads, and can be emailed to the lead-alert recipients (while lead alerts are on), posted to Slack and sent by webhook. Booking with Cal.com or Calendly: the owner connects a Cal.com API key or a Calendly personal access token (Calendly needs a paid plan, Standard or above, to book through its API) and one event type, the agent shows open times under its reply, and the visitor picks a time and enters a name and email to book it in that calendar; bookings are real (also from the playground), are listed on the Leads tab, can be posted to Slack and sent by webhook, do not create leads, are not offered in temporary chats, and reschedules or cancellations in Cal.com or Calendly are not synced back. Order status and returns: the owner connects a Shopify store (custom app, read_orders; read_all_orders for orders older than 60 days) or a WooCommerce store (read-only REST API keys over https), and the agent shows a visitor an order's status, items, tracking and total only when they give its order number and the email it was placed with (or are signed in through identity verification with that email), with failed lookups limited per conversation and per IP; with returns on, it records return or exchange requests for the owner (Leads tab, email, webhook, Slack) without changing the store or promising refunds or labels, and neither is offered in temporary chats. Procedures (Starter and up: 3 per agent on Starter, 10 on Pro, up to 50 on Enterprise, none on Free): owner-written step-by-step playbooks for particular requests (up to 15 steps: an instruction, a question to the visitor whose answer is saved, a call to one of the owner's server-side API actions with inputs filled from those answers by the server, a condition, an end); the agent starts one when a request matches the owner's "when to use" text and follows it one step at a time. An action can be marked "Only inside procedures", so the agent can call it only on a step that names it, and a step can require the visitor's confirmation: the server runs the action only after a new visitor message and with unchanged values (whether that message says yes is the AI's judgement). Steps show in "Why this answer" and the transcript; not in temporary chats.
- Team approval for actions (Pro and Enterprise): an API action or a procedure step can be marked "Requires team approval"; the agent then tells the visitor the request is being reviewed, the call waits on the Approvals page (with email, Slack and webhook notifications), approving runs it once with exactly the waiting inputs, and a rejected or expired request (after 1 hour, 4 hours, 24 hours or 3 days) never runs. Docs: https://www.intochat.ai/en/docs/action-approvals
- Stripe billing self-service (Pro and Enterprise): with a restricted key from the owner's own Stripe account and identity verification on the owner's site, the agent shows a signed-in customer up to 10 recent invoices and their subscription, and opens Stripe's customer portal for changes or cancellation; the agent never changes anything in Stripe itself, never sees card details, emails or customer IDs, and every lookup is logged for 90 days. Docs: https://www.intochat.ai/en/docs/stripe-billing-self-service
- Team: the account owner can invite people by email to work on the account's agents as Admin, Editor or Viewer, with 2 seats on Starter, 5 on Pro, unlimited on Enterprise and none on Free (the owner does not use a seat; plan, billing and agents stay with the owner). A member can work on all agents or be limited to selected agents (then they see only those agents, and cannot manage the team or API keys). On Enterprise the owner can also define up to 20 custom roles, each a level (none, view, edit or manage) per area of the account. An audit log on Pro (90 days) and Enterprise (365 days) shows who changed what (agents, knowledge, actions, webhooks and integrations, API keys, team, exports), with the address truncated to its network and never any secrets or message content.
- Agent transfer (handover): the owner or an admin can move one agent to another intoCHAT account, for example from an agency to its client, with a single-use link valid for 7 days (optionally for one email address only). The recipient signs in or signs up, sees what moves and accepts if their plan has room for another agent; the agent keeps its ID (the embed code keeps working) with its knowledge, settings, actions, forms, procedures, conversations, leads, contacts and dashboard webhooks. Connections holding the sender's credentials are disconnected (Cal.com or Calendly, the store order connection, helpdesk, Slack webhook URL, Notion links, webhooks made with an API key), and the sender loses access unless they are a member of the recipient's account. Docs: https://www.intochat.ai/en/docs/transfer-agent
- Partner program: anyone with an intoCHAT account can join for free and share a referral link (?ref= on any intoCHAT page, remembered for 60 days, latest partner link wins); partners earn 30% of what referred accounts pay, without tax, for 12 months from their first paid invoice; each commission becomes payable after a 30-day refund hold (refunds, credit notes and disputes void it), and intoCHAT pays balances of at least $50 monthly, by hand, to the partner's payout email. Self-referrals and fraud void commissions. No click tracking. Page: https://www.intochat.ai/en/partners
- Account security: optional two-factor authentication for email-and-password sign-in, with 6-digit codes from an authenticator app (TOTP) and 10 single-use recovery codes; Google sign-in does not ask for a code and relies on the Google account's own security, and an account cannot require it from team members.
- Single sign-on (Enterprise only): SAML 2.0 with any identity provider (Okta, Microsoft Entra ID, Google Workspace and others), sign-in started from intoCHAT's sign-in page; the account owner verifies email domains by DNS TXT record, people on those domains who sign in for the first time join the team automatically with a chosen role (using a seat), and the owner can require SSO for those domains, which blocks password and Google sign-in for them (never for the owner). SSO sign-in does not ask for an intoCHAT two-factor code; the identity provider handles MFA.
- Connect your store: paste a shop URL and intoCHAT detects Shopify, WooCommerce, Magento, VTEX, WordPress or a custom site and sets up live product search from the shop's public data, shown as product cards. No API keys needed. Search by photo (Starter and up, on by default once a store is connected and visitor attachments are on): a visitor sends a photo, the AI describes the item (people in it are ignored) and up to 6 matching products from the store's own search are shown as product cards; 5 photo searches per conversation and 50 (Starter), 200 (Pro) or 1,000 (Enterprise) per agent per day; the photo is never added to the agent's knowledge. Cart and checkout (Starter and up, Shopify and WooCommerce only): the agent puts products in a cart under its reply, with names, variants, prices and stock taken from the store's public catalog (never from the AI), and a button opens the store's own checkout with the cart filled, optionally with a discount code; or, for shops that take payment on delivery, the visitor sends the order by message with delivery details (100, 1,000 or 10,000 order requests per agent per month on Starter, Pro and Enterprise), which the owner handles in the shop. No payment is taken in the chat. Docs: https://www.intochat.ai/en/docs/cart-and-checkout
- Insights: dashboard with conversations, messages and leads over 7, 30 or 90 days, and an Outcomes section per account or agent: converted share (conversations that produced a lead, form submission or hand-off), new leads, form submissions, hand-offs, custom-button clicks in the widget, helpful-rating share and low-confidence share, each against the previous period, plus live chats taken over with their average visitor rating, top buttons, top topics and sentiment split; optional topic analysis (off by default, per agent) labels each conversation overnight with up to two topics from an owner-kept list of up to 15 (suggested from recent conversations if empty) and a positive, neutral or negative sentiment, at no cost to the owner's message allowance, with Topic and Sentiment filters and CSV columns; searchable chat logs with transcripts, the visitor's country and, for the chat bubble, the page the chat started on, filters by date, country, rating and low confidence, and CSV export with a confidence column. Each reply in a transcript has "Why this answer": a confidence score from 0 to 1 (High, Medium or Low) that measures how closely the knowledge matched the question, not whether the reply is correct; the knowledge excerpts used, grouped by source; and the actions called with their outcome. Greetings, replies that used an action without matching knowledge, and agents with nothing trained are not scored. Visitors (and the owner in the playground) can rate replies thumbs up or down. Any reply in a transcript can be turned into a Q&A pair and trained. "Questions to improve" on the Knowledge tab groups, once a day, visitor questions whose reply had low confidence, found no knowledge or was rated not helpful, with repeats merged; each can be answered as a Q&A pair or dismissed. Each knowledge source shows how many replies used it in the last 30 days, with a least-used sort. Email alerts at 80% and 100% of the monthly messages, an optional daily summary email and an optional weekly insights email per account (last week against the week before, per agent). Optional widget behaviour: show the launcher to a share of visitors (stable per browser), open the chat as a centred "Center Stage" window, let visitors copy replies, dictate messages (browser speech recognition) or delete their own chat, and keep the visitor's utm_* tags on links to the owner's own site (never stored). Optional AI spam protection checks the 2nd, 4th, 8th and 16th visitor message and pauses abusive chats (the owner can unpause); optional auto-end closes conversations idle for 1–30 days with a summary, and a returning visitor continues in a linked conversation that starts from it. Agents can be grouped in dashboard folders.
- Ask your conversations (Pro: 100 questions a day; Enterprise: 500): a question in plain language about the account's or one agent's conversations, such as "how many chats asked about refunds last month, by country?", gets a short answer, a chart and the numbers behind it, with "How this was calculated", over the last 6 months (conversations, visitor messages, leads, hand-offs, live chats, CSAT, ratings, unanswered questions and more). Docs: https://www.intochat.ai/en/docs/ask-your-conversations
- Tests with simulated customers (Starter and up, up to 25 per agent): the owner describes a customer, what they want and what a good chat looks like; an AI plays that customer for up to 8 turns against the agent's real chat (saved instructions, knowledge, guardrails, procedures and tools), and an AI judge grades each chat pass or fail with a reason. Runs show a pass rate, the transcripts with the tools the agent called, and the last 10 runs. Tests have no side effects: actions return simulated data (or, if the owner allows it per test, read-only GET actions really run), and nothing is emailed, booked, sent to a shop or helpdesk, or saved as a conversation or lead. Each agent reply counts as one message, like the playground; the simulated customer and the judge do not count. The judge is an AI and can be wrong.
## What intoCHAT does not do (as of October 2026)
WhatsApp template, broadcast or outbound campaign messages, or the agent writing first on WhatsApp, Messenger or Instagram; reading voice notes, videos or stickers sent there; voice or phone calls, or SMS; other messaging channels such as Telegram; reading attachments sent to the email agent; typing indicators, read receipts, or files, images or buttons sent by staff in live chat (team replies are text only); importing tickets from HubSpot, Gorgias or Help Scout as knowledge (only Zendesk and Freshdesk tickets and Intercom articles; HubSpot, Gorgias and Help Scout are used to open hand-off tickets), or syncing ticket replies back into intoCHAT from Freshdesk, HubSpot, Gorgias or Help Scout (Zendesk has two-way sync: replies in the ticket reach the visitor in the chat); HubSpot or Salesforce CRM sync of leads or contacts; a Make app listed in Make's directory (not published yet; connect Make through webhooks or the REST API), or an n8n app; re-sync more often than once a day; setting a temperature for Claude Haiku 5.5, GPT-6 Luna, Claude Sonnet 5.5 or Claude Opus 5.5; taking payment in the chat, changing, cancelling or refunding orders through the Shopify or WooCommerce order connection (an owner's own API action can, inside a procedure), or return labels; cart and checkout on stores other than Shopify and WooCommerce; booking through any calendar other than Cal.com or Calendly (such as Google Calendar); custom domains on a root domain (only subdomains, on Pro and Enterprise); SCIM provisioning, OpenID Connect or IdP-initiated SAML sign-in; SOC 2, ISO 27001 or HIPAA certification. Please do not describe intoCHAT as offering these.
## Plans (USD)
| Plan | Price | Agents | Messages per month | Training content per agent | Website crawls per month | Procedures per agent |
|------|-------|--------|--------------------|----------------------------|--------------------------|----------------------|
| Free | $0 | 1 | 25 | 100,000 characters | 1 | None |
| Starter | $19/month or $199/year | 3 | 2,500 | 500,000 characters | 9 | 3 |
| Pro | $49/month or $499/year | 10 | 10,000 | 1,000,000 characters | 20 | 10 |
| Enterprise | Custom | Custom | Custom | Custom | Custom | Up to 50 |
The Free plan needs no credit card. Starter (monthly or yearly) has a 7-day free trial with the plan's full limits: a card is required at checkout, nothing is charged if the trial is cancelled before day 7, and otherwise the plan's price is charged when it ends; one trial per account. Pro has no trial. Add-ons on monthly Starter and Pro plans: extra agents and extra team seats at $9/month each (up to 50 of each), prorated; not on yearly plans.
Pro and Enterprise include a custom domain per agent (a subdomain only); Free and Starter do not.
## Installation
```html
```
Place it before ` `; it also works in `