Reply to every WhatsApp message — even when you're not online.

It's a chatbot that sits on your WhatsApp Business number and replies to customers using a large language model (OpenAI, Gemini, Groq, etc.) instead of pre-written button flows. You write a personality prompt that describes your business and how the bot should sound. The bot uses that prompt plus the recent conversation to reply on its own.

The native WhatsApp auto-reply is one fixed message — "we'll get back to you soon". An AI chatbot actually understands what the customer asked and responds with a real answer. If you want the simpler, rule-based version, check our auto-reply rules feature — it's free and good for basic cases.

Yes — you bring your own AI key. We support enterprise-grade provider support (OpenAI, Anthropic, Google Gemini, Groq, and more). Most of them have generous free tiers, so you can start without paying anything for AI itself. If you don't have a key, we'll walk you through setting up a free one in 2 minutes.

Yes. Honestly, this is the bit we obsess over most. Customers in India type in Hindi, Hinglish, Gujarati, Marathi, Tamil — often switching mid-sentence. The bot handles that naturally because the underlying AI models (especially Gemini and the latest GPT/Claude versions) understand all of it without translation.

Honestly, it's only as good as the personality prompt you write (or generate). It works off that prompt plus the recent conversation context — it doesn't have a separate knowledge base. If your primary AI rate-limits or fails, ChatMitra automatically tries your backup providers (you can add up to 3). And if everything fails, the customer sees the fallback message you wrote during setup. We'd rather be honest about this than oversell it.

It depends on volume. Most businesses we work with are spending ₹20–30k/month on someone (or part of someone's time) handling WhatsApp queries. The chatbot handles the repetitive 60–75% — pricing, delivery, business hours, return policy — so that role either gets freed up for higher-value work, or you delay the next hire. See the pricing page for the actual numbers.

If you already have your WhatsApp Business API set up — under 5 minutes. If you're starting from scratch, the WhatsApp/Meta verification can take a day or two (that part isn't us, it's Meta). After that, the bot itself is quick.

On our servers in India, encrypted at rest. We don't share it with the AI providers beyond the message text needed to generate a reply. We're a Meta Business Partner, so we follow their data handling rules too.

A token is roughly 4 characters of English text — about three-quarters of a word. So 'How are you?' is about 4 tokens. AI providers like OpenAI and Gemini bill you per million tokens you send in (input) and receive back (output). For a typical WhatsApp conversation of ~17 messages, you use about 2,500 tokens total. That's why a 1,000-conversation month often costs only a few hundred rupees on cheaper models.

Context window is how much text the AI can read at once — the system prompt, your business info, and the recent conversation, all combined. It's measured in tokens. GPT-5 family has 400K tokens (about 300,000 words), Claude Sonnet 4.6 has 1M, Llama 3.3 has 128K. For WhatsApp chat, even 128K is plenty — most conversations are well under 5,000 tokens. The bigger windows matter for long B2B negotiations or businesses that want the bot to remember weeks of customer history.

Input tokens are everything you send TO the AI — your business personality prompt, the customer's message, recent conversation history. Output tokens are what the AI sends BACK — the bot's reply. Output is almost always more expensive per token than input. Example: GPT-5 Mini costs $0.125 per 1M input tokens and $1.00 per 1M output. That's why long bot replies cost more than long customer questions.

It's the instruction the AI reads BEFORE every customer message. The system prompt tells the bot who your business is, how to talk, what it can say (your prices, hours, return policy), and what NOT to say. ChatMitra writes this for you when you click 'Generate with AI' in Step 3 of the chatbot setup — you pick a tone and language, the system prompt is generated automatically. You can edit it later.

A reasoning model (also called a 'thinking' model) pauses to work through a problem step-by-step before replying — like OpenAI's o3, DeepSeek-R1, or Claude with extended thinking. They're better at multi-step problems (loan eligibility math, comparison logic, complex troubleshooting) but slower and 3-10× more expensive than regular chat models. For everyday WhatsApp chat, a regular model like GPT-5 Mini or Llama 3.3 is faster and cheaper.

Temperature controls how predictable vs creative the bot's replies are. Range is usually 0.0 to 2.0. At 0.0, the bot always picks the most likely next word — same answer every time, very robotic. At 2.0, it improvises a lot — friendlier but occasionally goes off-script. Most ChatMitra businesses set it to 0.7 (the default) — friendly but reliable. Crank it down to 0.3 for FAQ-style replies, leave at 0.7 for sales conversations.

A multimodal model can understand more than just text — usually images and sometimes audio. Gemini 2.5 Flash, GPT-5, and Claude Sonnet 4.6 are all multimodal. For WhatsApp chat, this means a customer can send a product photo, a damage picture, or a screenshot — and the bot can describe what it sees and reply about it. Useful for returns ('did this arrive damaged?'), retail ('do you have this style?'), and clinics ('is this rash serious?').

If your primary AI provider rate-limits you or has an outage, ChatMitra automatically tries the next provider in your fallback chain — and so on, up to 3 backups. Customers don't see an error. A typical setup: Groq as primary (fast and cheap), Gemini as backup #1, OpenAI as backup #2. If none of them work, the customer sees a fallback message you wrote (something like 'Our system is busy, please call us at...').

Rate limit is a cap on how many requests per minute (RPM) or tokens per minute (TPM) you can send to an AI provider. Free tiers have tight limits — Groq's free tier is 30 RPM and 6,000 TPM, which is fine for testing but breaks during a Diwali sale spike. When you hit a rate limit, requests fail until the next minute. ChatMitra's fallback chain catches these failures and routes to a backup provider so customers never see a broken reply.

RAG is when an AI looks up information from a document store before replying — instead of relying only on its training data. Useful for businesses with deep FAQ documentation (insurance policies, product manuals, course curricula). Cohere's Command R+ is built specifically for RAG. ChatMitra's AI Chatbot doesn't use a separate document store — it relies on the personality prompt and conversation context. For document-heavy use cases, you'd typically pair ChatMitra with a separate knowledge base.

It varies. Some providers (Gemini, Groq, Cerebras) have permanent free tiers with rate limits — useful for production at small scale. Others (OpenAI, Anthropic) give a one-time signup credit (~$5) and then charge per token. Some (Cohere) offer free Trial keys that are explicitly NON-COMMERCIAL — fine for testing only. ChatMitra shows the exact free-tier details on each provider's setup page so you don't get surprised when you go live.