Code Interpreter

CodeSandbox extends a regular sandbox with a stateful kernel — variables, imports, and function definitions persist between calls, just like a Jupyter notebook. It’s the right tool when an AI agent generates and runs code incrementally.

It boots from the code-interpreter-v1 template (Python 3.14 + Node.js 26, with a code-execution API on port 8080) and supports Python, JavaScript, and Bash out of the box.

Run code

import { CodeSandbox } from '@lizard-build/sdk';
 
// Every sandbox belongs to a project — usage is metered per project
const sandbox = await CodeSandbox.create({ project: 'my-project' });   // defaults to 'code-interpreter-v1'
 
await sandbox.runCode('x = 42');
const result = await sandbox.runCode('print(x * 2)');
console.log(result.stdout);        // "84\n"  — x survived from the previous call
 
await sandbox.kill();

runCode(code, opts?) returns an Execution:

Field	Description
`stdout` / `stderr`	Captured output streams
`results`	Rich results (values, and any generated images / charts)
`error`	An `ExecutionError` (`name`, `message`, `traceback`) if the code threw
`executionCount`	Monotonic counter for the kernel

Choose a language

Default is Python. Pass language for a one-off in another runtime:

const js = await sandbox.runCode('1 + 1', { language: 'javascript' });
console.log(js.results[0].data);   // "2"
 
await sandbox.runCode('echo "$(uname -s)"', { language: 'bash' });

Streaming output

For long-running cells, stream stdout/stderr as it’s produced instead of waiting for the result:

await sandbox.runCode('for i in range(5): print(i)', {
  onStdout: (line) => process.stdout.write(line),
  onStderr: (line) => process.stderr.write(line),
  onResult: (r)    => console.log('result:', r),
  onError:  (e)    => console.error('error:', e.name, e.message),
});

Isolated contexts

A context is an independent namespace inside the same sandbox — use one per agent session or per user, so their variables never collide.

const a = await sandbox.createContext({ language: 'python' });
const b = await sandbox.createContext({ language: 'python' });
 
await sandbox.runCode('secret = 1', { context: a });
await sandbox.runCode('print(secret)', { context: b });   // NameError — b never saw it
 
await sandbox.listContexts();
await sandbox.restartContext(a);    // clear all variables/state
await sandbox.deleteContext(b);     // free its resources

Pass either context or language to runCode, not both — a context already has a language bound to it.

Checkpointing a session

Because CodeSandbox is a sandbox, pause and resume work the same. Install a heavy dependency stack once, pause, and resume later with the kernel state intact:

const sandbox = await CodeSandbox.create({ project: 'my-project' });
await sandbox.runCode('import subprocess; subprocess.run(["pip", "install", "scikit-learn"])');
const id = sandbox.sandboxId;
await sandbox.pause();
 
// Later — resume with packages and kernel variables already in place
const resumed = await CodeSandbox.connect(id);
const out = await resumed.runCode('import sklearn; print(sklearn.__version__)');
console.log(out.stdout);
await resumed.kill();