The sandbox, the patching and the 3 a.m. pages are yours.
Judge0 and Piston are free to download. Running them safely is the expensive part.
Send code. Get a verdict. Code execution API for 46 languages, with judging built in.
Every run gets a fresh sandbox and returns output, exit code, time and memory, or a verdict for every test case. One online compiler API for online judges, coding courses, assessments and AI agents.
100,000 runs a month free/commercial use/no card
# Monthly sales in, totals and a chart out (pandas + matplotlib)
import pandas as pd
import matplotlib.pyplot as plt
units = list(map(int, input().split()))
months = ["Jan", "Feb", "Mar", "Apr", "May", "Jun"]
sales = pd.DataFrame({"month": months[:len(units)], "units": units})
print("Total units:", sales["units"].sum())
print("Best month:", sales.loc[sales["units"].idxmax(), "month"])
plt.figure(figsize=(4.5, 3))
plt.bar(sales["month"], sales["units"])
plt.title("Units sold per month")
plt.show()
Total units: 963 Best month: Jun
Python comes with pandas and Matplotlib, and charts come back with the output: Code interpreter API.
46 languages, on current versions. How long a small program takes, end to end:
Python 3.14 0.56s C GCC 15 0.67s C++ G++ 15 1.31s Java OpenJDK 25 2.25s C# .NET SDK 10 3.08s F# .NET SDK 10 4.61s Go 1.26 1.26s Rust 1.93 0.92s PHP 8.5 0.58s Ruby 4.0 0.69s Haskell GHC 9.12 1.63s TypeScript Deno 0.58s JavaScript Node.js 24 0.60s SQL SQLite 3 0.53s Bash 5.2 0.52s Fortran gfortran 15 0.69s Visual Basic .NET SDK 10 3.24s Kotlin 2.4 4.59s R 4.6 1.04s Swift 6.4 1.56s COBOL GnuCOBOL 3.2 1.05s Perl 5.44 0.73s Julia 1.13 1.11s Pascal Free Pascal 3.2 0.98s Octave 11 1.27s Assembly NASM 3 0.68s Lua 5.5 0.56s OCaml 5.4 1.40s Common Lisp SBCL 2.6 0.61s Erlang OTP 29 1.41s Elixir 1.20 1.69s Prolog SWI-Prolog 10 0.62s D LDC 1.43 2.13s FreeBASIC 1.10 1.04s Dart 3.13 1.42s Clojure 1.12 1.64s Groovy 6.0 3.03s Scala 3.10 3.68s Zig 0.17 3.99s Nim 2.2 2.05s Crystal 1.22 8.49s Ada GNAT 15 1.22s PowerShell 7.6 1.88s Racket 9.3 0.99s Objective-C Clang 21 1.81s
languages
46on current versions
free runs a month
100,000commercial use included, no card
tests in one request
24032 on the Free plan, one score
containers reused
0ever. One per run, then deleted.
Running other people’s code is a job nobody wants.
Old runtimes, results you poll for, prices on request.
Hosted APIs that stopped at Java 17 and Ruby 2.7.
Send the code. Get the answer in the same call.
No queue to run, no token to poll, no kernel to patch.
A verdict for every test. One request.
Send code and test cases to /api/run-tests/. Each test runs in its own sandbox. You get a verdict per test, an overall verdict and a score.
A stored problem, live: Judge runs its two samples, and Submit runs all five tests, three of them hidden, and returns the score with a signed receipt.
Best month
The first line holds the units sold each month. Print the month with the most units; if two tie, the earlier one.
hidden
Hidden tests
Come back with only their verdict, never the input.
×1
Compiled once
C, C++ and Java compile one build that every test runs.
240
Tests a request
32 on the Free plan, within a 60-second budget. One call, one score.
mcp
Async, signed, and for AI agents
Signed webhooks and receipts anyone can verify; run_tests is a tool on the MCP server.
Run tests referenceOnline judge API guideSigned receiptsVerifiable results for assessments
A dashboard for judging, too.
Write, test and publish stored problems in the dashboard, and see how each submission to them was judged, without writing code. It works on the same problems, submissions, sessions and webhooks as the API, so your team can use either.
- A problem editor
- Statement in Markdown with a preview as candidates see it, languages, checker, scoring, starter code and tests, typed or uploaded. Tests are hidden unless you make them samples.
- Test the draft
- Judge your own code against the draft before you publish it, and see each test's data and output, hidden ones included.
- Versions with a reference check
- When the problem has a reference solution, publishing runs it against every test first, and the version goes live only if it passes.
- Submissions, receipts and statistics
- Each submission's verdict, score, result per test, code and signed receipt, and each problem's acceptance rate, average score, verdicts and last 30 days.
- Widget sessions and webhooks
- Revoke a session or make a test one to try Submit. The webhook log shows every delivery and its attempts; resend one, or send a signed test webhook.
- Team members
- Invite up to 25 people by email, on every plan. Admins manage keys, widgets and the team; Members work on problems and submissions; only the Owner sees billing.
46 languages. Current versions.
Java 25, Go 1.26 and Ruby 4.0, not Java 17 and Ruby 2.7. Python ships with NumPy, pandas, SciPy, scikit-learn and Matplotlib.
- Python 3.14
- C GCC 15
- C++ G++ 15
- Java OpenJDK 25
- C# .NET SDK 10
- F# .NET SDK 10
- Go 1.26
- Rust 1.93
- PHP 8.5
- Ruby 4.0
- Haskell GHC 9.12
- TypeScript Deno
- JavaScript Node.js 24
- SQL SQLite 3
- DuckDB 1.5
- Bash 5.2
- Fortran gfortran 15
- Visual Basic .NET SDK 10
- Kotlin 2.4
- R 4.6
- Swift 6.4
- COBOL GnuCOBOL 3.2
- Perl 5.44
- Julia 1.13
- Pascal Free Pascal 3.2
- Octave 11
- Assembly NASM 3
- Lua 5.5
- OCaml 5.4
- Common Lisp SBCL 2.6
- Erlang OTP 29
- Elixir 1.20
- Prolog SWI-Prolog 10
- D LDC 1.43
- FreeBASIC 1.10
- Dart 3.13
- Clojure 1.12
- Groovy 6.0
- Scala 3.10
- Zig 0.17
- Nim 2.2
- Crystal 1.22
- Ada GNAT 15
- PowerShell 7.6
- Racket 9.3
- Objective-C Clang 21
46 languages
Median run time for a small program, compile included: 20 runs per language on production on 18 October 2026, sent from the API server, so your own network round trip comes on top.
All languagesTry them in the online compilerHow each one is compiled
From a server, a web page or an AI agent.
REST API: REST API reference
curl -X POST https://api.onlinecompiler.io/api/run-code-sync/ \
-H "Authorization: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"compiler": "python-3.14", "code": "print(\"Hello, World!\")"}'
# 200 OK
{
"output": "Hello, World!\n",
"error": "",
"status": "success",
"exit_code": 0,
"signal": null,
"truncated": false,
"time": "0.0248",
"memory": "8192",
"total": "0.0330"
}Widget and iframe: Embed the widget
<script src="https://onlinecompiler.io/widget.js"></script>
<div class="runcode" data-key="YOUR_WIDGET_KEY" data-lang="rust-1.93"></div>
<!-- Or, where only iframes are allowed -->
<iframe src="https://api.onlinecompiler.io/embed/?key=YOUR_WIDGET_KEY&lang=rust-1.93"
title="Runnable Rust example" width="100%" height="420" style="border: 0"></iframe>WebSocket: WebSocket guide
const socket = io("wss://api.onlinecompiler.io", {
auth: { token: "YOUR_KEY" },
});
socket.emit("runcode", { compiler: "go-1.26", code });
socket.on("codeoutput", (result) => render(result.output));MCP server: MCP code execution server
claude mcp add onlinecompiler \
https://api.onlinecompiler.io/mcp \
--transport http \
--header "Authorization: Bearer YOUR_KEY"- list_compilers46 ids, versions
- run_codestdout, exit, time
- run_testsverdicts, score
- judge_problemstored problems
Start from a template.
Three starter apps on GitHub, MIT licensed, that already call the API. Clone one, add your free API key and run it.
Online judge starter
A LeetCode-style practice site in Next.js: Run on the sample tests, Submit on the samples plus hidden tests, and a result for each test.
React code editor
A CodeMirror editor with a language menu, an input box and a Run button. A small route on your server adds the key, so it never reaches the browser.
Classroom autograder
An assignment template that grades student code on every push with GitHub Actions and puts the score in the job summary. For teachers leaving GitHub Classroom.
The Python and Node.js clients and the MCP server are on GitHub too, also MIT licensed.
Assume every submission is hostile. We do.
Each run starts in a fresh gVisor container. gVisor handles system calls itself, so user code never reaches the host kernel. When the run ends, the container is deleted.
- 06Your user’s code
Whatever they submit. Assume the worst.
- 05No network
Nothing in, nothing out. No installs, no phoning home.
- 04Non-root, no capabilities
An unprivileged user with every capability dropped.
- 03Read-only image
Nothing a run writes outlives it.
- 02gVisor
A user-space kernel answers every system call.
- 01Host kernel
Never talked to directly by user code.
- 2CPUs
- 512 MBmemory
- 16processes
- 30 sper run
lifecycleevery run
- createa fresh container: gVisor, read-only, offline
- runthe code, as a non-root user
- judgeits output against the test
- deletethe container is removed
How the sandbox worksAll limits, including code and output size
On Judge0 or Piston? Change the base URL.
Your client code stays. Swap the URL and the key header, and keep shipping.
- 66/71
- Judge0 CE language ids run, older ones on the current version
- 46/46
- languages answer to Piston names and aliases
- wait=true
- supported on the Judge0 API, so no polling loop
Judge0 migration guideMoving off PistonComing from Sphere Engine
const client = new Judge0({-Remove: baseUrl: "https://judge0-ce.p.rapidapi.com/submissions",+Add: baseUrl: "https://api.onlinecompiler.io/api/judge0/submissions",-Remove: header: "X-RapidAPI-Key: YOUR_RAPIDAPI_KEY",+Add: header: "X-Auth-Token: YOUR_API_KEY", });
+2 −2 · that’s the whole migration
Side by side with Judge0 and Sphere Engine.
We lose two rows. We left them in.
| usOnlineCompiler.io | Judge0 | Sphere Engine | |
|---|---|---|---|
| Free plan | 100,000 runs a month, commercial use included | None at checkout | 14-day trial |
| Paid from | $10 a month for 150,000 runs | $44.99 a month for 2,000 runs a day (RapidAPI) | Quote only |
| Judging | Per-test verdicts, hidden tests and a score in one request | One expected output per submission, batches of 20 | Stored problems and custom judges, in the separate Problems module |
| Getting output | One call | Submit, then poll a token | Create, poll every 2 s, fetch each stream |
| Newest versions | Java 25 · Go 1.26 · Ruby 4.0 | Java 17 · Go 1.23 · Ruby 2.7 | Java 22 · Go 1.22 · Ruby 3.1 |
| Time per run | 30 s, compile included | 5 s CPU by default, up to 20 | 1–15 s, compile excluded |
| Memorythey win | 512 MB | 256 MB by default, up to 2 GB | 1.5 GB |
| Sandbox | gVisor, fresh per run, no network | isolate | Not disclosed |
| AI agents | Official hosted MCP server | Community MCP wrappers | None |
| Self-hostingthey win | Hosted only | Yes, GPL-3.0 | On request |
From each product's public documentation and pricing pages, checked on 18 October 2026.
Judge0 alternatives, with JDoodle, OneCompiler and CodapiSphere Engine alternative
100,000 runs a month. Free.
Commercial use included, no card. Paid plans stop at their limit, so there is no overage bill to dread.
The Free plan covers it. $0, commercial use included.
fits you
Free
$0/mo
- 100,000 runs a month
- 4 runs at once
- 32 tests a request
- Commercial use, no card
fits you
Starter
$10/mo
- 150,000 runs a month
- 8 runs at once
- 60 tests a request
- $0.067 per 1,000 runs
fits you
Pro
$49/mo
- 1,000,000 runs a month
- 10 runs at once
- 120 tests a request
- $0.049 per 1,000 runs
fits you
Businesssoon
$199/mo
- 4,000,000 runs a month
- 20 runs at once
- 240 tests a request
- $0.050 per 1,000 runs
More than 4,000,000 runs, or a contract with your name on it? Ask about Enterprise
Questions developers ask.
Anything else? Talk to a human
Yes. POST the source code, a compiler id and optional input to https://api.onlinecompiler.io/api/run-code-sync/ and the response holds stdout, stderr or compiler errors, the exit code, run time and memory. It covers 46 languages, from Python, Java and C++ to Rust, Go and SQL.
Yes. The Free plan includes 100,000 runs a month in every language, commercial use included, with no credit card. Paid plans start at $10 a month for 150,000 runs, and Pro works out to $0.049 per 1,000 runs.
Yes. POST /api/run-tests/ runs one program against up to 240 tests (32 on Free) and returns a verdict for each, such as accepted, wrong_answer or time_limit_exceeded, plus an overall verdict and a score. Hidden tests return only their verdict, and C, C++ and Java compile once per request.
Yes. OnlineCompiler.io serves Judge0's API at https://api.onlinecompiler.io/api/judge0, so a Judge0 client moves by changing its base URL and key, with nothing to install and no cgroup v1 or privileged containers. 66 of Judge0 CE's 71 language ids run, and wait=true works.
Since 15 February 2026 emkc.org only serves whitelisted non-commercial educational projects. OnlineCompiler.io serves Piston's /runtimes and /execute at https://api.onlinecompiler.io/api/v2/piston, so a Piston client needs a new base URL and a free API key.
Each run gets a fresh container under gVisor, which handles the program's system calls in user space, so the code never talks to the host kernel directly. It runs as a non-root user on a read-only image with no capabilities and no network, capped at 512 MB of memory, 2 CPUs, 16 processes and 30 seconds, and the container is deleted when the run ends.
Yes, through the MCP server at https://api.onlinecompiler.io/mcp. It gives Claude, Cursor, VS Code and other MCP clients four tools: list_compilers, run_code, run_tests and judge_problem. The agent's code runs in the sandbox instead of on your machine, and each agent can have its own key with a monthly limit.
In Python, yes: NumPy, pandas, SciPy, scikit-learn, Matplotlib, seaborn, statsmodels, SymPy, openpyxl, Pillow and DuckDB, at published versions, and charts come back with the result. Other languages have their standard library, and programs have no network access to install more packages. If your problems need other packages, tell us which ones.
Ship the Run button this afternoon.
Get a key, send a POST, read the verdict. The first 100,000 runs every month are on us.
$ claude mcp add onlinecompiler https://api.onlinecompiler.io/mcp --transport http --header "Authorization: Bearer YOUR_KEY"