What is a code execution tool?

A code execution tool is a tool an LLM calls to run the code it writes, returning the output so the model can use it in its answer.

You can host the tool yourself in a Runtime sandbox, for any model, with the internet on your terms and a bill per second of real use. A new sandbox ran its first Python command 351 ms after the create request at the median on 24 September 2026 (speed), fast enough to start one per call.

Why it matters for AI agents

A model predicts text; it does not compute. Asked to sum a column or parse a log, it guesses. Given a tool that runs code, it writes a short program, reads the real result and answers from that. The same tool lets an agent test a patch, convert a file or draw a chart.

Tools come in two kinds, and the difference decides who runs the code:

A server tool runs on the model provider's side. You add it to the request and the provider runs the code in its own container. Anthropic's code execution tool is one: "The API runs every command server-side and returns the results to Claude within the same request".
A client tool is one you define. The model returns a call, your code runs it wherever you choose, and you send the result back. A sandbox is the usual place, because the code came from a model.

Hosted tools in facts

Tool	Where the code runs	Documented limits
Anthropic code execution tool	Anthropic's container, 1 CPU, 5 GiB RAM, 5 GiB disk	Internet "completely disabled"; pre-installed packages only
Anthropic pricing	1,550 free hours a month per organization	Then $0.05 an hour per container, 5-minute minimum
OpenAI Code Interpreter tool	"A fully sandboxed virtual machine"	Python; memory 1g (default), 4g, 16g or 64g
OpenAI container lifetime	Expires after 20 minutes unused	Later calls to it fail
Your own tool on Runtime	A Firecracker microVM per sandbox	Size, packages, network and lifetime set per sandbox

Anthropic's code execution is free when the same request also includes its web search or web fetch tool.

Build one as a client tool

Give the model a tool such as run_code(command) and back it with a sandbox. withruntime/tools hands you the tool already written, with a name, a description, a JSON Schema and a function:

TypeScriptimport { Sandbox } from "withruntime";import { sandboxTools } from "withruntime/tools";await using sbx = await Sandbox.create({ network: { internet: false } });const [exec] = sandboxTools(sbx);console.log(exec.name, exec.description);console.log(await exec.execute({ command: "python3 -c 'print(6 * 7)'" }));

Pythonfrom withruntime import Sandboxfrom withruntime.tools import sandbox_toolswith Sandbox.create(network={"internet": False}) as sbx:    run, read, write, ls = sandbox_tools(sbx)    print(run("python3 -c 'print(6 * 7)'"))

Pass the schema to any model's tool format, run execute when the model calls it, and return the output. Adapters for the OpenAI Agents SDK, the Vercel AI SDK, the Claude Agent SDK, LangChain and others wrap the same tools (frameworks).

Runtime's side of the tool

Any model, since the sandbox is yours and the tool is ordinary code.
Network rules that allow a package index, deny everything, or switch the internet off, enforced on the host.
Any package or image, with pip, npm and sudo apt-get working.
Cost: $0.025 per vCPU-hour of measured CPU and $0.0075 per GiB-hour of memory, with no minimum (pricing).

For notebook-style sessions that keep variables and return charts, see code interpreter.

Sources

Checked 25 September 2026.

Anthropic: code execution tool: how it runs, container limits and pricing
OpenAI: Code Interpreter: container, memory options and expiry

Facts on this page were checked on 25 September 2026.