Build with a coding agent

Connect the Tavus MCP server so your agent creates the deployment and builds the integration, or paste an embed or widget prompt into an agent without MCP.

There are two ways to hand the work to a coding agent.

• Connect the Tavus MCP server when your agent supports MCP, as Cursor, Claude Code and Codex do. The agent creates the deployment, reads the integration prompt itself and builds the integration. Nothing to copy.
• Paste a prompt when it does not, as in Lovable, Replit, Bolt and v0. You create the deployment in the PAL Maker and paste the embed or widget prompt into the chat.

Connect the Tavus MCP server

1. Add the server

   Point your client at the hosted endpoint and sign in once in the browser. For Codex:

   codex mcp add tavus --url https://mcp.tavus.io/mcp
   

   For Claude Code:

   claude mcp add -s user --transport http tavus https://mcp.tavus.io/mcp
   

   Other clients and the sign-in flow are on the MCP server page.

2. Ask for a deployment

   Tell the agent which PAL to put on your site and how, for example “Deploy my support PAL as a widget on this site” or “Put the onboarding PAL in a card on the pricing page”. It checks for a deployment it can reuse, asks where the site is hosted, how much use to allow and whether each visitor keeps their own memory, and creates the deployment once you confirm. The deployment is live as soon as it exists.

3. Let it build

   The deployment comes back with its id, the embed or widget prompt to follow and the PAL’s page tools, so the agent builds the integration in your stack without asking for any of them and writes a handler for every tool the PAL can call on the page. Later, tavus_deployment_get returns the current tool list; settings changes and taking a deployment down happen in the PAL Maker.

The three tools are listed under Deployments in the MCP tools reference.

Choose a prompt

Embed prompt

A conversation inside your page: a section, a card that opens a modal, a full-screen flow, or a kiosk.

Open prompt

Widget prompt

A floating launcher: use the default button, add your own trigger, or let the PAL act on your site.

Open prompt

Open a prompt on GitHub, click Copy raw file above the file, and paste the full text into your agent’s chat. Both prompts work in Cursor, Claude Code, Codex, Lovable, Replit, Bolt and v0.

1. Create a deployment

   In the PAL Maker, create a deployment with the Embed or Widget channel and copy its id. The id is safe to expose; access is controlled by the deployment’s allowed origins, call limits and bot protection.

2. Paste the prompt

   Copy the full text of your chosen prompt into the agent’s chat. Tell it what you want to build, such as a card that opens a modal or a “Talk to us” button in the header.

3. Answer its questions

   Before writing code the agent reads the reference pages and asks what the visitor should experience, where it goes in your app and which stack to keep, your deployment id, and whether the PAL has tools for acting on the page. Plain words are enough.

4. Test on a published URL

   The call needs camera and microphone access. Previews that run inside an iframe may not grant it. If a preview stalls on permissions, publish and test there, and add that domain to the deployment’s allowed origins if you have restricted them.

Need inspiration first? Browse the six embed examples. For platform-specific setup, see App builders and website builders.

Agents can read this documentation directly. The index is at docs.tavus.io/llms.txt, and every page is available as markdown by appending .md to its URL. The prompts link to those .md URLs.

Next steps

Examples

Six complete embed pages to read or copy before you prompt.

App builders and website builders

Lovable, Replit, Bolt and v0 with a prompt; Webflow, Framer, Wix and Squarespace with a script tag.