Build and run long-running AI agents on secure MicroVMs.
Bring any agent framework, or build one with our Velos AI agent. Corevelos runs each agent persistently in its own hardware-isolated MicroVM, scales from one to tens of thousands, and bills only for what you use.
Prompt
Build a habit-tracker Android app and ship an APK.
An illustration of how Velos AI orchestrates agents and tools across Corevelos sandboxes.
Four ways to create and manage sandboxes.
Use Corevelos however you work. Start in the web app with no install, script it from the CLI, build it into your product with the TypeScript or Python SDK, or let your agents create their own over MCP.
Web app
no installCreate and manage sandboxes from the browser. Nothing to set up, instant to start.
app.corevelos.com
CLI
your shellDrive the full sandbox lifecycle from your terminal.
corevelos sandbox create
SDK
TypeScript & PythonBuild sandboxes into your product with first-class SDKs in both languages.
await Sandbox.create()
MCP
for AI agentsExpose sandboxes over MCP so Claude, Cursor, and your agents create them directly.
corevelos mcp
Migrating from E2B? Swap your import for @corevelos/sdk/e2b and keep the rest of your code. SDK and CLI reference.
Start from a prebuilt template.
Boot a sandbox with a popular agent or tool already installed, or bring your own as a custom image. Every template runs in its own hardware-isolated MicroVM.
Open-source coding agent, ready to run.
Autonomous software-engineering agent.
Anthropic's agentic coding CLI.
OpenAI's coding agent.
Pair-programming agent for your repo.
General-purpose autonomous agent.
Open agent runtime, prebuilt and ready.
Personal autonomous agent.
An agent that drives a full desktop.
Headless browser for web automation.
Build from a Dockerfile or with an AI agent.
Meet Velos, your master AI agent.
Velos AI is a canvas-based master agent. Give it a goal and it plans the work, spawns the right agents, and orchestrates them across Corevelos sandboxes, so you can stand up a whole agent system from a single prompt.
- Describe a goal; Velos breaks it down and assigns each step to an agent.
- Every agent it spawns runs in its own hardware-isolated Corevelos sandbox.
- Watch and steer the entire run on a live canvas.
Persistent agents that resume in about a second.
Each agent runs in its own MicroVM that boots in about 128 ms and can freeze to disk with its memory intact, then resume right where it stopped. Run a single agent or scale to tens of thousands.
Persistent by default
Pause an agent to disk with its memory intact and resume right where it stopped, instead of rebuilding context on every run.
Snapshot, fork, and park
Fork a running agent and the child inherits its live memory for parallel work. Idle agents cost almost nothing, so you can keep thousands parked without burning compute.
The runtime that keeps your agents alive.
Agents do not finish in a single request. They run for hours, pause for input, and resume for days. Corevelos is the runtime underneath: each agent lives in its own hardware-isolated MicroVM that persists, survives restarts, and stays contained, so it runs for as long as the work takes.
- Runs as long as the work takes
- Agents run for hours or days. Pause one to disk when it is idle and resume it with memory intact, so a long task picks up exactly where it left off.
- Durable by design
- An agent can sit paused for weeks and come back to the same filesystem and memory. State survives restarts, so nothing is lost between runs.
- Hardware isolation
- Each agent runs in its own MicroVM with its own kernel, filesystem, and network, so untrusted agent code stays contained to a single sandbox.
- Contained and capped
- A per-sandbox egress policy and hard spend caps keep a runaway or prompt-injected agent from exfiltrating data or running up the bill.
What teams build on Corevelos.
If your product runs code it generated, drives a browser, or spins up environments on demand, it belongs in a sandbox.
Coding agents
A real shell, filesystem, and PTY to edit, run, and test code in a disposable environment.
Code interpreter and data analysis
Run model-generated Python against user data without exposing your own backend.
Computer-use and browser agents
A browser or desktop environment per agent for web automation.
Evals and reinforcement learning
Fork hundreds of identical environments in parallel, run them, and tear them down.
Long-running autonomous agents
Agents that work for hours or days, pausing and resuming without losing state.
One organization, one bill, every project with its own limits.
Group agents into projects inside your organization. The org pays a single bill while each project keeps its own limits, keys, and cost.
- Projects
- Isolate resources, keys, and spend; organizations roll them into a single bill.
- Attribution
- Per-key usage attribution shows exactly what each project and credential consumed.
- Audit logs
- Every action is recorded, so you can see who created, paused, or killed what.
- Spend caps
- Account-level caps stop a runaway bill before it happens.
Only pay for what runs.
Pay as you go, billed by the second. Start with $200 in free credits, no credit card and no subscription. Every feature is included from your first free sandbox.
Questions, answered.
What you need to know about running agents on Corevelos. The docs go deeper.
How is this different from running containers?
A container shares the host kernel; a MicroVM gives each agent its own kernel, filesystem, and network. For running untrusted, agent-generated code, that boundary is the point.
What am I billed for?
The vCPU, RAM, and storage an agent holds, billed by the second it runs against your prepaid balance. Paused agents bill storage only, and killed ones bill nothing.
Can I bring my own code or framework?
Yes. Bring any agent framework, start from a prebuilt template, convert a Docker image into a template, or have Velos AI build one for you. Every feature is included from your first free sandbox.
I already use E2B. Can I switch?
Swap your E2B import for @corevelos/sdk/e2b and keep the rest of your code.
Is my data isolated from other customers?
Yes. Every agent runs in its own MicroVM with its own kernel, filesystem, and network. Nothing is shared between sandboxes.
Give your agents a place to run.
Create your first sandbox in under a minute. $200 in free credits, no credit card required.