agent0 interacts with the world using one universal tool to rule them all: code.

the concrete tool is exec_ts and it can be used like this:

• agent0 (x-ai/grok-code-fast-1) writes code in typescript on the fly to provide an answer to what ever prompt is given
• the tool exec_ts runs the generated code in a e2b sandbox & provides the result back to the agent
• agent0 can self-heal on errors, as it just fixes the code and runs exec_ts again
• we define the context like apis to use (for example: use google rss for news)

this is a foundation. take it, adapt it, make it yours.

inspired by theo's video on Cloudflare's "code mode".

1. install dependencies

pnpm install

2. get API keys and configure

cp .env.example .env

edit .env and add your keys:

required:

• E2B_API_KEY - create account at https://e2b.dev (free tier: 500 executions/month)
• OPENROUTER_API_KEY - create account at https://openrouter.ai (pay per use, ~$0.10/1M tokens for grok-code-fast-1)

optional:

• BRAVE_API_KEY - create account at https://brave.com/search/api (free: 2000 queries/month)

3. start dev server

pnpm dev

launches Mastra with agent0

4. chat with agent0

open playground at http://localhost:4111, select agent0, start chatting.

try these prompts:

• "give me the top 5 news from germany today and the weather in berlin"

(adapt these)

we include these as examples to show you how it works. replace them with APIs your agent needs:

• Google News RSS - https://news.google.com/rss (free, no key)
• Brave Search - https://brave.com/search/api (requires BRAVE_API_KEY)
• Open-Meteo Weather - https://open-meteo.com (free, no key)

to add your own: open src/mastra/agents/agent0.ts, scroll to the API section in the prompt, add:

### Your API Name
**Endpoint**: https://your-api.com/endpoint?param={value}
**Auth**: Header "Authorization: Bearer {YOUR_API_KEY}" (if needed)
**Returns**: { data: { field1, field2 } }

agent reads this, understands the API, writes code to use it. no code changes needed on our side. so provide the stuff that you think is the most important.

(optional)

if you have domain-specific helpers that the agent uses repeatedly (data schemas, validators, parsers), mount them instead of having the agent rewrite them every time:

1. create utils/validators.ts:

export function validateEmail(email: string): boolean {
  return /^[^\s@]+@[^\s@]+\.[^\s@]+$/.test(email);
}

2. mount when calling tool:

import fs from "fs";
import { execTsTool } from "./mastra/tools/exec_ts.js";

const validators = fs.readFileSync("utils/validators.ts", "utf-8");

await execTsTool.execute({
  context: {
    code: `
      import { validateEmail } from "./validators.ts";
      console.log(JSON.stringify({ 
        ok: true, 
        data: { valid: validateEmail("[email protected]") }
      }));
    `,
    files: { "validators.ts": validators },
  },
});

saves tokens, keeps logic consistent across calls.

this is your foundation. build on it.

want different APIs? edit the prompt.
want different model? change one line.
want custom utilities? mount your files.