> ## Documentation Index
> Fetch the complete documentation index at: https://tinytalk.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Frequently Asked Questions
> Answers to the most frequently asked questions about Tiny Talk.
These answers are based on the most common questions from our users. If you don't find what you're looking for, contact us at [support@tinytalk.ai](mailto:support@tinytalk.ai).
## General
Tiny Talk is an AI agent platform that lets you create agents trained on your own content. You can embed them on your website, connect them to WhatsApp and Slack, and manage conversations through a built-in Help Desk.
No. You can create an agent, upload content, and embed the widget on your site without any coding. The embed code is a simple script tag you paste into your website's HTML.
Yes. The free plan includes 1 agent, 25 AI credits and 25 messages per month, 600K characters of knowledge base storage, and access to GPT-3.5, GPT-4o, and GPT-4o Mini. No credit card required.
Tiny Talk is a web-based platform. There is no native mobile app, but the dashboard works in mobile browsers. The chat widget is fully responsive on mobile devices for your visitors.
## Security & privacy
Tiny Talk's infrastructure is hosted in the European Union. Core services run on **AWS** (Frankfurt), **Hetzner** (Germany/Finland), **Supabase** (Germany), and **MongoDB Atlas** (Germany). AI model requests are sent to providers like OpenAI (US) or OpenRouter (US) unless you use the Azure EU option on the **Pro** plan, which keeps AI processing within European data centers.
Yes. All data is encrypted **in transit** using TLS (HTTPS) and **at rest** using AES-256 encryption across all storage systems — including uploaded documents in S3, databases, and backups.
Inference provider API keys (OpenAI, OpenRouter, Groq) are encrypted at rest and only decrypted at request time. Your Tiny Talk API secret key is hashed with a one-way function — even Tiny Talk cannot reverse it.
Yes. Files uploaded to the Knowledge Base are stored in encrypted S3 storage and are not publicly accessible. You can only access them through the dashboard using cryptographically signed URLs that are generated at request time.
Key sub-processors include **AWS** and **Hetzner** (infrastructure), **Supabase** and **MongoDB Atlas** (databases), **OpenAI** and **OpenRouter** (AI model inference), and **Stripe** (payments). For the full list, see our [Privacy Policy](https://tinytalk.ai/legal/privacy-policy).
Tiny Talk takes data privacy seriously. Infrastructure is hosted in the EU, data is encrypted in transit and at rest, and tenant data is logically isolated. For EU data residency requirements, the **Pro** plan offers Azure EU models that keep AI processing within European infrastructure. A **Data Processing Addendum (DPA)** is available upon request for customers with compliance requirements — contact us at [support@tinytalk.ai](mailto:support@tinytalk.ai).
Yes. You can delete resources, conversations, and contacts through the dashboard at any time. For full account data deletion, contact [support@tinytalk.ai](mailto:support@tinytalk.ai).
## Agent & conversations
The agent uses AI models (like GPT-5 or Claude) combined with your Knowledge Base content. When a visitor asks a question, the agent searches your uploaded documents and URLs for relevant information, then generates a response using the AI model.
No. The agent does not automatically learn from conversations. It only uses the content you've added to the Knowledge Base. To improve responses, update your Knowledge Base with new or better content.
Yes. Use the **System Prompt** in your agent settings to define the agent's personality, tone, language, and behavior. For example: "Be formal and concise" or "Use a friendly, casual tone."
You can instruct the agent in the system prompt to ask clarifying questions or collect information. For example: "Before answering, ask the user for their order number." The agent follows natural conversation patterns based on your instructions.
Yes. Use the [Help Desk](/features/help-desk) to monitor conversations and jump in at any time. When a human responds, the agent stops auto-replying for that conversation.
Yes. Visitors can request human assistance (takeover request). Your team receives notifications via the Help Desk and push notifications (if configured).
Suggested messages are pre-written prompts that appear as clickable buttons in the chat widget. They help guide visitors on what to ask. Configure them in [Messenger → Suggested Messages](/features/messenger#suggested-messages).
The most common cause is a missing inference provider API key. If your plan requires you to bring your own key (BYOK), you must add at least an **OpenAI API key** under **Integrations → Hub → OpenAI** before the agent can generate responses. Other things to check:
* **Message or credit limit reached** — If you've exceeded your monthly limit, the agent stops responding until the next billing cycle. Check your usage in **Playground** or purchase a [message credits add-on](/plans#add-ons).
* **Invalid API key** — Use the **Verify** button in the Integrations hub to confirm your key is working and has the required models and permissions. See [OpenAI's error codes](https://developers.openai.com/api/docs/guides/error-codes) for common API key issues.
* **Knowledge Base not trained** — If you've added content but haven't trained the agent, it may not have context to respond. Check that your resources show a **Trained** status in the Knowledge Base.
## Knowledge base & training
PDF, Word (.docx), Excel (.xlsx), PowerPoint (.pptx), CSV, Markdown (.md), MDX (.mdx), and plain text (.txt). Files must contain selectable text — images embedded in documents are not processed. See [Knowledge Base](/features/knowledge-base) for the full list.
Yes. Use the [website crawler](/features/website-crawler) to automatically scrape your site's content. Go to **Knowledge Base → Add Resource** and enter a URL.
Common causes: the file exceeds your plan's size limit, the file is password-protected (PDFs), or it contains only images without selectable text. See [Knowledge Base troubleshooting](/features/knowledge-base#troubleshooting).
**Characters** measure the total size of your training data (documents, URLs, etc.) — this is a permanent limit that frees up when content is deleted. **Messages** count visitor messages and human agent replies — this resets each billing cycle. **Credits** are consumed when the AI generates a response, with the cost depending on the model. See [Plans](/plans#understanding-your-limits) for details.
Crawling is processed in a shared queue. During peak times, there may be delays. If a crawl doesn't start within 30 minutes, try submitting the URL again. Very large sites (1,000+ pages) take longer to process.
Yes. Connect either service from **Integrations** in the dashboard, then import resources from **Knowledge Base → Add Resource**. See [Knowledge Base](/features/knowledge-base) for setup steps.
## Tools
Tools let your agent take action during a conversation — call your API, book a meeting, look up an order, file a support ticket, or trigger a workflow — instead of just answering from the Knowledge Base. There are two types: **Custom Tools** (HTTP endpoints you define yourself) and **Platform Tools** (first-party integrations Tiny Talk maintains for you, including Calendly and Cal.com for scheduling). The agent decides when to call them based on the descriptions you write. See [Tools](/features/tools).
The Knowledge Base is for *information* — static text the agent searches to answer questions ("What's your refund policy?"). Tools are for *actions* and *live data* — the agent calls an endpoint at conversation time to fetch a fresh answer or change something in another system ("What's the status of order ORD-12345?"). Most agents benefit from both.
Tools are available on all paid plans. Basic AI gets 5 tools per agent, Standard AI gets 10 per agent, and Pro AI gets 15 per agent. The Free plan does not include Tools. See [Plans](/plans) for the full breakdown.
Yes. Secrets you save on a tool are encrypted at rest with AES-256 and decrypted only at the moment a tool is invoked. The dashboard shows only a masked preview after saving — the raw value is never sent back to the browser. Authorization headers are redacted from logs.
No. The tool runtime calls from public infrastructure and blocks IP literals, private and loopback addresses, and cloud-metadata endpoints.
The model picks tools based on the **Name**, **Description**, and **When to use** fields. If it skips your tool, those descriptions usually need to be more specific about the kinds of messages it should match. Make sure the tool is enabled, and that the AI model you've selected supports tool calling. See the [troubleshooting section](/features/tools/custom-tool#troubleshooting) for more.
No. Tool calls don't consume extra credits — credit cost is determined by the AI model's response, the same as for non-tool conversations.
Yes. **Calendly** and **Cal.com** are available as Platform Tools — install either (or both) from **Agent → Tools → Platform Tools**, paste your event-type URLs, and the agent offers the right one when a visitor wants to book. On the website widget the booking flow is embedded inline; on WhatsApp and Slack the agent posts a personalised link. Cal.com supports `cal.eu` URLs if you need EU-hosted scheduling. See [Calendly](/features/tools/calendly) and [Cal.com](/features/tools/cal-com). More platform tools are on the way.
No. Calendly and Cal.com event-type URLs are public, so you only paste the URLs into the configuration — Tiny Talk never authenticates against your scheduler account. Free and paid Calendly or Cal.com accounts both work.
## AI models
It depends on your plan. **Credit-based plans** (Basic AI, Standard AI, Pro AI) include AI credits — no API key needed. **Legacy BYOK plans** require you to provide your own inference provider API key (e.g., OpenAI, OpenRouter, or Groq).
For most use cases, **GPT-4o Mini** or **GPT-5.2 Chat** offer the best balance of quality and cost (1 credit). For complex queries, **GPT-5.2**, **Claude Sonnet 4.6**, or **Gemini 3.1 Pro** (2 credits) are excellent. Reasoning models (like GPT-5 Series) are powerful but can be slow to respond and consume more credits — test them thoroughly before using in customer-facing scenarios to ensure the use case justifies the added cost and latency. See the full [AI Models](/features/ai-models) guide.
Yes. Tiny Talk supports 50+ models across OpenAI, Anthropic (Claude), Google (Gemini), Meta (Llama), xAI (Grok), DeepSeek, Mistral, Cohere, and Qwen. All are available on paid plans.
Credits are consumed each time your agent sends a response. Most models cost 1 credit, while premium models cost between 2–10 credits. Credits reset monthly. See [Plans](/plans#understanding-your-limits).
Reasoning models spend extra compute "thinking" before they respond, which improves quality on complex questions at the cost of higher latency and credit usage. Tune the trade-off in **Agent Settings → Model Configuration** with the **Reasoning Effort** control. See [Reasoning configuration](/features/agent-settings#reasoning-configuration) and the [AI Models](/features/ai-models#reasoning-capable-models) page for the list of reasoning-capable models.
## Messenger (Widget) & Installation
Copy the embed script from your agent settings and paste it before the closing `` tag on your website. See the [installation guide](/install) for platform-specific instructions (WordPress, Shopify, Wix, etc.).