On Runtime, every sandbox has one, in seven languages, inside its own Firecracker microVM. Python, JavaScript, TypeScript, R, Java, Bash and Go run as notebook-style sessions.
Why AI agents use one
A model is poor at arithmetic and data work done in its head. It is good at writing code that does that work. A code interpreter closes the loop. The model writes a cell, the interpreter runs it, and the model reads the output and writes the next cell. Because state persists, the model can load a file once and then ask ten questions about it, the way a person works in a notebook.
Running that code requires isolation. The model may receive a hostile file, or a prompt injection may tell it to read your secrets. An interpreter inside a sandbox limits the damage to a disposable machine.
Runtime's code interpreter
| Fact | Value |
|---|---|
| Languages | Python, JavaScript, TypeScript, R, Java, Bash, Go |
| In the image | Python, JavaScript, TypeScript and Bash; R, Java and Go install on first use |
| State | Variables persist between runs; Go keeps functions, types and imports |
| Rich results | matplotlib charts and R plots as PNG; pandas and R data frames as tables |
| Files | display returns a .png, .svg, .html or .csv file as a result, in any language |
| Python data tools | NumPy, pandas and matplotlib in the default image |
| Isolation | A Firecracker microVM with its own kernel; network rules enforced on the host |
On 23 September 2026, the first use of Java or Go took about 35 seconds. The
first use of R took about 90 seconds. This happens once per sandbox. To skip the
wait, install them in a custom image with apt
(code interpreter).
Run a cell
TypeScriptimport { Sandbox } from "withruntime";await using sbx = await Sandbox.create({ network: { internet: false } });await sbx.interpreter.run("import math\nx = math.pi");const cell = await sbx.interpreter.run("round(x * 2, 3)");console.log(cell.results[0]?.data["text/plain"]); // 6.283Pythonfrom withruntime import Runtimeruntime = Runtime()with runtime.sandboxes.create() as sbx: sbx.interpreter.run("import math\nx = math.pi") cell = sbx.interpreter.run("round(x * 2, 3)") print(cell["results"][0]["data"]["text/plain"])The second cell reads x from the first. To run another language, pass
{ language: "typescript" }, or language="r" in Python. An agent connected
through MCP gets the same sessions through the runtime_sandboxes_interpreter_run
tool (MCP).
What it costs
The interpreter adds no cost beyond its sandbox. Runtime charges $0.025 per vCPU-hour of measured CPU and $0.0075 per GiB-hour of memory. A 2 vCPU, 4 GiB sandbox costs $0.03125 an hour while the model thinks, and $0.08 an hour with both CPUs busy (pricing). A new account gets 100 free hours of a 1 GB machine in its first 30 days, without a card.
Related
- A code interpreter for chatbots
- A data analysis agent
- How to run untrusted code from an LLM safely
- What is an agent sandbox?
Facts on this page were checked on 25 October 2026.