Vercel AI SDK tools

npm install onlinecompiler ai

The tools need ai version 5 or later. Set ONLINECOMPILER_API_KEY to an API key from the dashboard (free, no card), then hand the tools to the model:

import { generateText, stepCountIs } from 'ai';
import { onlinecompilerTools } from 'onlinecompiler/ai-sdk';

const { text } = await generateText({
  model: 'openai/gpt-5-mini',
  tools: onlinecompilerTools(), // reads ONLINECOMPILER_API_KEY
  stopWhen: stepCountIs(5),
  prompt: 'What is the 40th Fibonacci number? Compute it, don\'t guess.',
});
console.log(text);

The tools

• runCode takes compiler, code and optional input, and runs the program as /api/run-code-sync/ does: up to 30 seconds and 512 MB, no network. Exit code 124 means it ran out of time.
• runTests takes compiler, code and tests (each an input and an expected_output), with an optional checker (trim, exact, tokens, float, rows), tolerance and time limit, and judges as /api/run-tests/ does: an overall verdict, a score and a verdict per test.
• onlinecompilerTools() gives both, sharing one client. Each function takes the client's options (apiKey, baseUrl, timeout, fetch) or a client you made.

import { runTests } from 'onlinecompiler/ai-sdk';

const tools = {
  // Only the judging tool, with the key passed in instead of read from the environment.
  runTests: runTests({ apiKey: process.env.MY_ONLINECOMPILER_KEY }),
};

What the model gets back

runCode returns the run's result, with any files the program wrote listed by name and link, not their contents:

{
  "status": "success",
  "exit_code": 0,
  "output": "102334155\n",
  "error": "",
  "time": "0.0213",
  "memory": "9216",
  "truncated": false,
  "files": []
}

runTests returns verdict, score, passed, total, compile_error, a plain-words summary, each test's verdict, output and time, and the submission id, under which the full result and its signed receipt stay available. An error from the API (a bad key, a limit reached) reaches the model as a failed tool call.

The same client covers the rest of the API: see the Node.js client. Agents in Claude, ChatGPT, Cursor and VS Code can use the MCP server instead.