OnlineCompiler MCP Server

An MCP server that lets AI assistants execute code in 46 programming languages via OnlineCompiler.io.

Each run starts a fresh gVisor sandbox (512 MB RAM, 30 s timeout, no network access).

Supported Languages

Python, C, C++, Fortran, Java, C#, F#, Visual Basic, Go, Rust, PHP, Ruby, Haskell, TypeScript (Deno), JavaScript (Node.js), Kotlin, R, Bash, SQL (SQLite), SQL (DuckDB), Swift, COBOL, Perl, Julia, Pascal, Octave, Assembly (NASM), Lua, OCaml, Common Lisp, Erlang, Elixir, Prolog, D, FreeBASIC, Dart, Clojure, Groovy, Scala, Zig, Nim, Crystal, Ada, PowerShell, Racket and Objective-C. list_compilers returns the current list with versions.

Setup

Point your MCP client at the URL. No install needed.

https://api.onlinecompiler.io/mcp

There are two ways in:

• Sign in (OAuth). Clients that support MCP sign-in (claude.ai, Claude Desktop and mobile, ChatGPT, Claude Code, Codex) open an OnlineCompiler.io sign-in page (Google or an email link) and ask you to approve. The app then appears as a key named after it, for example "Claude (connected app)", on your API Keys page. Disable that key to disconnect the app.
• API key. Send your key as Authorization: Bearer <key> (the bare key works too). Get a free key on the API Keys page.

Runs count against your plan either way. A wrong or disabled key gets 401.

Tools

list_compilers

List all available compilers with their identifiers.

run_code

Execute code with a specified compiler.

Parameters:

• code (required): Source code to execute
• compiler (required): Compiler identifier from list_compilers (e.g. python-3.14, gcc-15)
• input (optional): Standard input for the program
• time_limit (optional): Seconds the run may take, above 0 and at most 30 (default 30)
• files (optional): Files the program can use, up to 20 and 1,000,000 decoded bytes in total: helper source files (compiled with the program) and input data. Each {"name": "util.py" or "pkg/util.py", "content": "...", "encoding": "utf8" or "base64"}; the program finds them in its working directory

python-3.14 includes numpy, pandas, scipy, scikit-learn, matplotlib, seaborn, statsmodels, sympy, openpyxl, pillow and duckdb. Files a program writes to $OUTPUT_DIR come back (png, svg, csv, json, txt; up to 8 files and 2 MB), and plt.show() saves each open figure as a PNG.

Returns: output, error, status, exit_code, time, memory, truncated, files, files_truncated. PNG files arrive as MCP image content; SVG, CSV, JSON and text files are listed with their text (cut at 64 KB).

run_tests

Check a program against test cases and judge it like an online judge. At most 10 tests per call, each counting as one run. Every test runs unless the code fails to compile or the request runs out of time.

Parameters:

• code (required): Source code to check
• compiler (required): Compiler identifier from list_compilers
• tests (required): Up to 10 objects, each {"input": "...", "expected_output": "..."}
• time_limit (optional): Seconds per test, above 0 and at most 30 (default 2); compiling does not count, and Java, Groovy and F# get a start-up allowance on top
• checker (optional): trim (default), exact, tokens, float or rows (unordered lines, numbers within tolerance)
• files (optional): Files shared by every test (input data or helper source files), same shape as in run_code

Returns: an overall verdict, a score from 0 to 100 and a verdict per test.

judge_problem

Judge code against one of your stored problems. Hidden tests return only their verdict, never their input or expected output. Nothing is invented: a missing problem or a failed run is an error (for example HTTP 404: Problem not found).

Parameters:

• problem (required): The id or slug of one of your stored problems
• compiler (required): Compiler identifier from list_compilers
• code (required): Source code to judge
• mode (optional): all (default) runs every test; samples runs only the sample tests

Returns: a verdict, a score, a per-test verdict, a summary, a submission id, a signed receipt, and the problem id and version.

Client Configuration

claude.ai, Claude Desktop and mobile

In claude.ai or Claude Desktop, open Customize, then Connectors, then Add custom connector. Paste https://api.onlinecompiler.io/mcp, sign in to OnlineCompiler.io when asked and approve. The connector belongs to your Claude account, so it works in the Claude mobile apps too.

ChatGPT

Turn on developer mode in ChatGPT's settings, add a custom connector with https://api.onlinecompiler.io/mcp and sign in when asked.

Claude Code

claude mcp add --transport http onlinecompiler https://api.onlinecompiler.io/mcp

Then run /mcp inside Claude Code, pick onlinecompiler and sign in. To use an API key instead:

claude mcp add onlinecompiler https://api.onlinecompiler.io/mcp --transport http --header "Authorization: Bearer YOUR_API_KEY"

Codex

codex mcp add onlinecompiler --url https://api.onlinecompiler.io/mcp
codex mcp login onlinecompiler

To use an API key instead, skip the login and put the key in the environment variable that bearer_token_env_var names, in ~/.codex/config.toml:

[mcp_servers.onlinecompiler]
url = "https://api.onlinecompiler.io/mcp"
bearer_token_env_var = "ONLINECOMPILER_API_KEY"

VS Code (Copilot / Claude extension)

Add to .vscode/mcp.json:

{
  "servers": {
    "onlinecompiler": {
      "url": "https://api.onlinecompiler.io/mcp",
      "headers": {
        "Authorization": "Bearer your-api-key"
      }
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "onlinecompiler": {
      "url": "https://api.onlinecompiler.io/mcp",
      "headers": {
        "Authorization": "Bearer your-api-key"
      }
    }
  }
}