How to Build with ElevenLabs and Cursor at a Hackathon

Make Your Demo Talk Back

This is the loop I run to add a voice to a hackathon project. ElevenLabs generates the speech, and Cursor's agent writes the code around it. I write a spec, connect the agent to ElevenLabs, then let it build and test one feature at a time while I babysit the spend.

Step 0: Write the PRD Before Cursor Opens a File

The agent writes the code, but it can't guess what you want. Before Cursor touches a file, write down what you're building, who it's for, and what done looks like. Save it as PRD.md and AGENTS.md at the repo root so every prompt reads from the same source.

Skip the PRD and the agent keeps reinventing your data model and adding features you never asked for, burning API credits on every wrong guess. With it, the agent can type the CRUD endpoints while you decide what the interface looks like and whether the output is any good. Spend the hours you save on UX polish and the pitch.

Step 1: Check for an MCP First

An MCP (Model Context Protocol) server gives your agent a menu of typed tools for a service. Instead of reading pasted docs, the agent sees real actions like 'generate speech' or 'clone a voice' and gets structured data back.

ElevenLabs ships an official one at github.com/elevenlabs/elevenlabs-mcp. Add this to Cursor's MCP config and the agent can call text-to-speech, voice cloning, sound effects, transcription, and the ElevenAgents platform directly.

{
  "mcpServers": {
    "ElevenLabs": {
      "command": "uvx",
      "args": ["elevenlabs-mcp"],
      "env": {
        "ELEVENLABS_API_KEY": "<your-api-key-here>"
      }
    }
  }
}

Step 2: No MCP? Paste the Docs URL

Some tools don't have an MCP server yet. For those, put a link to the official docs inside the prompt. Cursor fetches the page and works from the live API instead of the stale version the model remembers from training.

Here's the shape of a prompt I use to build a streaming text-to-speech endpoint without the MCP. The 'Done when' section tells Cursor when to stop, so it doesn't wander off.

Read https://elevenlabs.io/docs/overview/intro and PRD.md > Backend Routes.

Create app/api/speak/route.ts in Next.js 15 (App Router, TypeScript).

Behavior:
- POST handler, accepts { text: string } JSON body
- Calls ElevenLabs text-to-speech with model_id "eleven_flash_v2_5"
- Voice ID and API key come from process.env (already in .env.local)
- Returns audio as a streaming audio/mpeg response

Constraints:
- Use the official @elevenlabs/elevenlabs-js SDK (install if needed)
- No retries, no caching, no logging
- Do not touch any other route or component

Done when:
- curl -X POST localhost:3000/api/speak -d '{"text":"hello"}' returns audio bytes
- The response Content-Type is audio/mpeg
- npm run build passes with no new TS errors

Step 3: Key In, Then Build and Test

Put your API key in .env.local at the project root, and keep that file out of git. Your app code reads the key from there. The MCP server gets its copy from the env block in mcp.json above.

ELEVENLABS_API_KEY=sk_your_key_here
ELEVENLABS_VOICE_ID=JBFqnCBsd6RMkjVDRZzb
ELEVENLABS_MODEL_ID=eleven_flash_v2_5

Now hand the agent one clean task: read the PRD, build the next unfinished feature, run the test, and stop only when it passes. You watch the loop and step in when it goes sideways.

Make the Agent Write the Test First

The most useful habit I've found: the agent writes the test first, then builds the feature that makes it pass. That's test-driven development (en.wikipedia.org/wiki/Test-driven_development) with an LLM writing the code.

“Write new code only if you first have a failing automated test.”

A test gives the agent a hard pass/fail signal, so it converges instead of guessing. You also end up with a regression suite, which matters when you swap voices or models five minutes before the demo.

Babysit the Spend, Then Ship

An agent in a tight loop spends real money on your Cursor subscription and your ElevenLabs credits. Runaway retries, endless tool calls, or the same audio regenerated over and over can burn through a free tier, so don't walk away from it at hour 18.

When you pitch it, don't spend your slot explaining the tools. When I judged LA Hacks, a couple of teams spent around 30 seconds explaining what OpenAI Agents or ElevenLabs Agents are. If the judge has used them, that time is gone, so ask whether they're familiar with ElevenLabs before you explain it.