Give OpenClaw
real compute, pay per second.
Register one tool in OpenClaw and your agents can spin up persistent sandbox environments, deploy full projects, and expose services publicly — without an account, without infrastructure configuration. Billed per second of active sandbox runtime.
Real OpenClaw agents that need live compute.
Each pattern below is an OpenClaw agent with the compute tool registered once. Each uses a different sandbox property — short-lived isolation, stateful pipelines, public deployment, or tiered compute.
Execute generated code safely inside the agent.
An OpenClaw agent that generates code needs somewhere safe to run it. The agent creates a sandbox as a tool call, executes the code via the process API, reads results back via the filesystem API, and tears down. Per-second billing means the sandbox charges only for the execution window — not the time the agent spends reasoning before or after.
Deploy a project and return a live URL.
An OpenClaw agent can take a project archive, upload it to a sandbox, and get back a public preview URL — all in a single chain of tool calls. The gateway handles dependency installation and process startup; the agent just receives the URL and passes it to the user or downstream steps. No DNS, no reverse proxy, no infrastructure work inside the agent.
Run a full build inside one environment.
An OpenClaw build agent checks out code, installs dependencies, runs tests, and reads the artifact — all inside one sandbox that stays alive through every tool call. No reinstalling packages between stages, no state lost between steps. Tear it down when the build finishes; the agent spins up a fresh sandbox for the next run. Pick a tier sized to the build.
Match sandbox tier to the task at hand.
An OpenClaw orchestration agent can spin up different tiers for different subtasks — a lightweight sandbox for file parsing, a larger one for model inference or compilation. Each sandbox is billed at its tier's per-second rate. OpenClaw budget caps keep total compute spend predictable across all sandbox interactions in a run.
One tool. Real compute inside every agent.
Sandboxes are persistent environments your OpenClaw agents interact with via process, filesystem, and network APIs. Deploy code by uploading files — JSON or tarball — and the gateway handles dependency installation and process startup. OpenClaw budget caps bound total sandbox spend; per-second billing means idle runs cost zero.
- Single OpenClaw tool
- Persistent sandboxes
- Per-second billing
- Budget caps honored
OpenClaw-specific questions.
If something below doesn't cover your case, ping us — we work directly with OpenClaw builders, no SDR funnel.
How does this register as an OpenClaw tool?
+
It's a set of POST endpoints — create sandbox, interact via process/filesystem/network, deploy, teardown. Register the base endpoint in OpenClaw as an HTTP tool with per-second pricing. OpenClaw uses the rate to enforce budget caps and to show the user what each sandbox interaction will cost before the agent runs.
Does OpenClaw need a separate account?
+
No. The gateway manages the underlying infrastructure entirely. OpenClaw pays per second from a wallet you connect — no account to create, no API key to store in agent configuration.
How do sandbox lifetimes work inside an OpenClaw agent?
+
The agent controls the sandbox lifetime explicitly — create it when compute is needed, use it across as many tool calls as required, and tear it down when done. Billing runs only while the sandbox is active. An agent that creates and tears down within a few seconds pays for only those seconds.
Can an OpenClaw agent deploy and expose a public URL?
+
Yes. Upload project files via the deploy tool call — the gateway installs dependencies and starts the process. Then call the preview endpoint to get a public URL. The whole sequence is OpenClaw tool calls; the agent receives the URL and can pass it to the user or downstream steps.
What happens to sandbox state between OpenClaw tool calls?
+
It persists. The sandbox keeps its process state, installed packages, and filesystem contents between calls. A file written in one tool call is readable in the next. This is what makes multi-step agents efficient — no re-setup, no data shuttling.
How do per-second costs interact with OpenClaw budget caps?
+
They stack. The sandbox accrues per-second charges while active. OpenClaw run budget caps bound total spend across all tool calls — compute included. If the cap is reached mid-task, the agent stops issuing new calls; you can tear down the sandbox and request more budget or hand off the partial result.