Runloop

How Runloop works

Runloop provisions microVMs it calls devboxes, aimed at coding agents at scale. It runs 10,000-plus in parallel, holds SOC 2, and cold start is around 400ms. Persistence works like git: disk snapshots you can branch from, so an agent can fork a known state and try several paths. It supports MCP, integrates with the OpenAI Agents SDK, and ships SWE-bench and Terminal-Bench harnesses so you can evaluate agent runs in the same place they execute. It fits teams building coding agents that need heavy parallelism and built-in evaluation.

Pricing in practice & watch-outs

Runloop sits at the enterprise end. The Pro plan starts at $250 a month before usage, and the per-core rate of $0.108 per vCPU-hour is toward the higher side of the category, so this is not the place to run casual experiments. You are paying for the parallelism, SOC 2, and the eval harnesses. It is also built specifically around coding agents, so if your workload is general code execution rather than SWE-bench-style agent development, a lot of what you are buying goes unused. Match it to the job before committing to the plan floor.