Learn how to configure settings and customize your chat widget
Core settings that control how your agent answers. They apply to every widget of the agent. Manage these on the Configuration page.
The predefined instruction or context sent to the underlying Language Model (LLM) when a user interacts with your agent. This prompt helps guide the LLM's responses and ensures the chat aligns with your intended behavior. For example: 'You are a helpful assistant for customer support' or 'Provide concise answers about our services.' This prompt will override any default prompts.
Under the prompt, click Check prompt to have AI grade it Weak, Fair, Good or Strong on four dimensions: Clarity, Boundaries, Grounding and Voice. Each dimension lists what is missing and why. Improve with AI suggests a rewrite and shows exactly what changed, and Fix next to a dimension limits the rewrite to that dimension. Use this prompt puts the suggestion in the editor — it still needs Save changes.
Select the AI model that the chat will use to answer user questions. You can choose a specific model from the available options, or select 'Auto' (default) where the system automatically analyzes each question and selects the most suitable model based on its complexity.
Controls how varied and creative the agent's replies are. Pick one of three presets — Precise, Balanced (default, recommended), or Creative — or use the slider to set any value between 0.00 and 1.00. Lower values stay closer to your knowledge base and produce stable, repeatable answers ideal for support, policy, and factual Q&A. Higher values introduce more variation in wording and tone, which is better for brainstorming, marketing copy, or conversational engagement assistants. The setting applies uniformly across every supported model (OpenAI and Anthropic), so switching models won't change the tone of your agent — switching style will.
When enabled, search results retrieved from your knowledge base are re-scored using a cross-encoder model before being sent to the LLM. This typically produces more relevant and precise responses at the cost of a small amount of additional latency per query.
Visual and behavioral settings that control how a widget looks and feels. Each widget has its own — open it under Integrations → Widget and use the Appearance tab. When you create another widget, you can copy the look of an existing one.
The title displayed at the top of the chat widget. This can be used to provide context for the chat, such as 'Customer Support', 'Ask Us Anything', or your company name.
The primary color used for the chat widget, including buttons and headers. Use a hex color code (e.g., #123456) to match your brand's theme and create visual consistency.
The name displayed for the chat agent, such as 'John Doe', 'Support Bot', or 'AI Assistant'. Enable 'Allow multiple names' to provide several names instead of one — the widget then picks one at random for each visitor's chat session and keeps it consistent throughout the conversation, making the agent feel more personal across users.
The icon for the chat widget. You can upload a PNG or JPEG image (square, 64–128px recommended, up to 150 KB). This icon will be displayed in the chat interface to represent your brand.
Show a small circle with the agent's logo next to each of its messages, as in common chat apps. This option is available only after an icon is uploaded; without a logo in use it stays disabled.
The text displayed on the chat button. For example, 'Start Chat', 'Ask a Question', or 'Get Help'.
The placeholder text shown in the search input field. This guides users on what they can search for, such as 'Search FAQs', 'Type your question', or 'What can we help with?'
Enable streaming responses to chat. When enabled, the chat widget will stream responses in real-time as they are generated, providing a more interactive and engaging experience.
The direction of the text in the chat widget. Options include: Auto (automatically detects based on language), Left-to-Right (LTR), or Right-to-Left (RTL) for languages like Arabic or Hebrew.
A widget puts your agent on a website. Manage your widgets under Integrations → Widget.
Click Create widget and enter your website's address. An agent can have several widgets — one per site, each with its own look and embed code — all answering from the same prompt, knowledge and tools. How many widgets an agent can have depends on your plan.
A widget only loads on the addresses it allows. The address you create it with is allowed right away; add others, such as a staging site, in the widget's Settings tab. Addresses are matched by scheme and host, so https://example.com and https://www.example.com are separate entries.
Copy the script from the widget's Install tab and paste it into your website HTML. The script names the widget by its id and sets its position. Embed codes copied before widgets existed keep working.
Pausing a widget hides it on every page that uses its embed code until you resume it. Deleting a widget stops its embed code for good; your agent and its knowledge are not affected.
The Dashboard highlights the setup steps left — adding a widget, allowing your website, adding knowledge, and installing the widget — so new users can finish deployment from one place.
Connect your own tools so your agent can act on live data, not just answer from documents. Manage these under Tools → MCP Servers.
Register a Model Context Protocol (MCP) server your team already runs — ticket search, order lookup, an internal knowledge API — by giving it a name and URL. Your agent becomes an MCP client and can call your tools live during a conversation, retrieving real-time or transactional information instead of answering only from uploaded documents.
After adding a server, click Scan to discover the tools it advertises along with their descriptions. Enable the specific tools you want your agent to use — disabled tools are never offered to the model. Re-scan at any time to pick up new tools; your enable and display choices are preserved by tool name.
Decide whether end users see tool activity. Turn on Show in chat to display a small indicator while a tool runs, and set a friendly Display text such as 'Searching ticket history…'. Leave it off to keep tool calls behind the scenes.
Servers can be open (no auth) or protected with a Bearer token. Tokens are stored securely in a secrets vault and are never displayed again after saving.
The server must speak MCP over HTTPS (streamable HTTP or SSE). Tool calls are bounded per conversation turn to keep responses fast, and tool output is treated as untrusted data. No website changes are needed — every widget of the agent automatically gains access to the tools you enable. Tools work with both OpenAI and Anthropic models.