Use casesCode interpreter
A code interpreter is a sandbox your application calls: send code, get back the output, the errors and any files. Each run is short and thrown away, so the price of a second of compute, how quickly a sandbox starts and how many can run at once matter more than anything else.
“My app lets users or an LLM run code, and I need somewhere safe to run it.”
This is infrastructure you build into a product, not something your users log into. Your server asks the provider's API for a sandbox, runs the code, collects the result and throws the sandbox away.
The code is untrusted by definition, so isolation is the first question: whether each run gets its own kernel, and what it can reach on the network. The isolation and outbound network pages compare that.
From its own documentation, each line linked to the page it comes from
Small. Provider minimums run from 0.125 of a core and 128 MiB on Modal upwards. Battleships' interpreter workload uses 1 vCPU / 2 GiB / 5 GiB and fills 80% of that memory, so 2 vCPU / 4 GiB / 10 GiB is the safer size.
No. Each run lasts seconds to minutes; 20,000 one-minute runs a month is about 650 a day.
battleships.dev, built by the Boat team
Usually nothing. Results go back to your app or your own storage.
battleships.dev, built by the Boat team
About 17% CPU on average in Battleships' interpreter workload. Charging by the second, and only for CPU actually used, both matter here.
battleships.dev, built by the Boat team
Easiest first. “Without a terminal” means someone who does not type commands can get it running from the provider's website.
| Provider | How | Without a terminal? | What it involves |
|---|---|---|---|
| Cloudflare Sandbox | Official guide or template | No | Sandboxes started from a Worker, with CPU billed only while working.Cloudflare Sandbox SDK ↗ |
| Daytona | Official guide or template | No | Sandboxes started from an SDK, billed per second, with $200 of free compute.Daytona docs ↗ |
| E2B | Official guide or template | No | Its original use: a code-interpreter SDK and per-second billing.E2B docs ↗ |
| Modal | Official guide or template | No | Sandboxes from Python or JavaScript, sized down to a fraction of a core.Modal: sandboxes ↗ |
| Tensorlake | Official guide or template | No | Wrap `sandbox.run()` as a `run_code` tool and keep one sandbox per session so installed packages and written files carry between calls; the Tool Calls page has Claude, OpenAI and OpenAI Agents SDK examples. Pre-install the usual libraries in a snapshot so each session starts warm, and turn egress off or to an allowlist before running model-written code.Tensorlake: tool calls ↗ |
| Vercel Sandbox | Official guide or template | No | Firecracker sandboxes from an SDK, with CPU billed only while working.Vercel Sandbox ↗ |
Read from each provider's documentation on 29 October 2026. A provider not listed as one click can still be the right home; it means someone has to follow its instructions.
Three starting points for different people, each with its catch
Built for exactly this. At the workload below it is about $55 a month at list price, and the free plan's one-time $100 credit covers most of the first two months.
Catch: The free plan limits sessions to an hour and disk to 10 GiB. Concurrency limits are per plan.
CPU is billed only while working, so the same workload is about $13 a month.
Catch: Smaller company, one US region, and a free tier capped at four sandboxes at once.
Started from a Worker next to your code, with CPU billed only while working.
Catch: The smallest size that has 2 vCPU also has 8 GiB of memory, which you pay for. There is a $5 monthly account minimum.
Beyond these providers. Anthropic's hosted code-execution tool runs Claude's code for you with no sandbox to manage: 1 CPU, 5 GiB, no internet, 1,550 hours included and then $0.05 per container-hour.
Compare these three head to head: features and cost side by side.
20,000 one-minute runs a month on 2 vCPU / 4 GiB / 10 GiB machines, thrown away after each run; results go to your own storage. The CPU works about 17% of the time. Change any number; open a row for the arithmetic.
List prices only, excluding model tokens, egressData the machine sends out to the internet, such as downloads to a user or uploads to another service. Some providers charge per GB; most agents send very little., taxes and anything a provider does not publish. How the numbers are made.
Each page compares every provider, with a source for every row
Isolation: what kind of computer you get — Every provider says “sandbox”. Underneath, it is a container, a gVisor container, a microVM or a full virtual machine, and that decides what your agent can install, run and break.
Outbound network: egress, allowlists and static IP — What the agent can reach, what can be blocked, and whether the outside world sees a stable address. This is where an agent that browses meets a firewall that was designed for a web app.
Ways to drive it — Dashboard, API, SDK, CLI, MCP, SSH, desktop. A provider can have all seven and still be unusable without a terminal, or have three and be genuinely click-driven.
Billing models — The meter decides your bill more than the hourly rate does. Four shapes are in play: flat per size, per vCPU and per GB, a plan fee that returns as credit, and measured CPU.