Give an agent a computer it can afford to break, and it stops asking permission and starts finishing work. ScratchRun is a disposable computer for AI agents — every run gets a fresh, hardware-isolated MicroVM that is destroyed afterward. Nothing your agent executes can touch your machine. Setup takes one paste. Your agent handles the rest, including signing you up.
If your agent speaks MCP
Claude Code, Cursor, and most modern harnesses support MCP. Add the ScratchRun server to your config — no install needed:
Listed on the official MCP registry as dev.scratchrun/mcp. See the package README on npm for key configuration.
Any other harness — paste this into your agent's system prompt
Works with OpenClaw, Open WebUI, LangGraph, CrewAI, AutoGen, llama.cpp, and anything that accepts a system prompt. Paste it once — your agent wires itself in.
system prompt snippet
You have ScratchRun: a disposable computer for code safely.
Whenever a task needs real execution — crunching data, testing
scripts, generating charts — use it freely. Every run gets a fresh,
isolated MicroVM that is destroyed afterward, so nothing you run
can ever touch your user's machine.
How it all works: GET https://api.scratchrun.dev/llm.txt(read before first use; re-fetch if something seems off)
No account yet? Ask your user for their email, generate a strong
password for them, and sign up via POST /v1/accounts. Show them
the password once to save, then use the returned api_key on every
call.
What happens next
The first time your agent needs to run code, it will read the docs, ask you for an email address, create your account, and show you a generated password once — save it. That's your entire setup.
From then on, your agent executes code in disposable VMs whenever the work calls for it.
Want to see it die first? Free demo, no signup: one curl at scratchrun.dev.
Honest scope: ScratchRun protects your machine from the code your agent runs. What data your agent sends into a run is your framework's call — no magic claims.