Free vs paid code execution servers you're pricing isolation, not CPU.

Free execution servers differ mostly in how much they trust the code. Shell Exec runs bash on the host directly — maximal power, zero containment, appropriate only when the agent and machine are both yours. podman-mcp-server interposes containers via Podman or Docker. runno wraps execution in a purpose-built sandbox, and RLM Tools keeps a persistent Python session for codebase exploration. Same category, very different failure modes.

All of them are free to run because they execute on compute you already have. The cost is the engineering judgment of matching containment level to how much you trust what the agent writes.

Real isolation has real overhead: keeping container runtimes patched, setting CPU, memory, and disk quotas so a runaway loop can't take the host down, scrubbing network access, and resetting state between runs so one execution can't poison the next. For multi-tenant setups — several agents, or agents acting for users you don't know — the bar rises to VM-level isolation, and the maintenance becomes a standing job.

Special hardware raises it further. mcp-server-colab-exec exists because GPU execution is its own provisioning problem: T4s and L4s aren't sitting under most desks.

A hosted execution endpoint sells the finished article: an isolated environment that exists for one run and is destroyed after. With x402 pricing, an agent buys it in-band — 402 Payment Required, price paid in USDC on Base, settlement in about two seconds, Ed25519-signed receipt, and the code runs only after payment clears. From $0.01 per call, with no account or API key to provision.

Per-execution is the honest unit here because executions are spiky and unequal: a hundred quick syntax checks one day, one GPU-hungry batch job the next. Paying per run prices each correctly; a subscription prices for the peak and idles the rest of the month.

Decide on trust first. Code written by your own agent against your own repo can run in a free local sandbox like runno or a container via podman-mcp-server. Code from agents or users you don't control needs isolation strong enough that you should ask whether you want to operate it at all — that's the paid endpoint's case. Then let volume tune the answer: constant heavy execution may justify owning infrastructure; intermittent execution almost never does.

If you already operate hardened sandboxes or GPU runners, that capability is sellable per run. Claim your listing on Loomal and price above your per-execution cost; the 5% fee on settled transactions is currently waived.