Agents need their own machine

Give your agents aReal
Computer

Celesto runs AI agents in secure sandboxes, observes how they behave, and helps them improve from every run.

Start free View docs

Backed by

Computer using agents love Celesto

Terminal

$ celesto computer create --cloud$ celesto computer exec COMPUTER_ID --cloud -- echo "Hello World"$ celesto computer ssh COMPUTER_ID --cloud

Celesto CLI

Your cloud computer, in three commands.

Create a computer, execute commands, and connect over SSH from your terminal.

celesto.ai/docs

Celesto DocsPreview

browser

Browser use

Give your agent a browser to navigate the web.

AgentFinding the browser setup guide

Browser use

A browser your agent can use.

Let your agent navigate websites, find information, and take action in its own cloud computer.

Human in the loopPreview

You’re in control

AgentYou

Help your agent with the next step

Hand back to agent

Agent paused · ready when you are

Human control

Step in. Then hand it back.

Take control of the browser when your agent needs help, then let it continue.

research-workspacePreview

Your workspace, ready again

research-notes.mdSaved

Browser sessionRestored

Project filesReady

Continue where you left off

Persistent sessions

Pick up where your agent left off.

Keep files, browser logins, and workspace state between runs, so your agent can continue without starting over.

Open-source core. Self-host or use our cloud.

Celesto

Apache-2.0

The microVM that boots every Celesto sandbox. Run it on your own infrastructure — no account, no lock-in.

$ celesto computer create

GitHub239stars

Celesto Cloud

Managed

Same CLI, same SDK, zero infrastructure. Sandboxes boot in seconds and you pay only while they run.

$ celesto computer create --cloud

Start free380developers building

Core features

Build, run, and improve AI agents.

Agents that improve themselves

Define an agent, connect tools and MCP servers, and deploy. The harness runs outside the sandbox and continuously learns from every run, so your agents become more reliable over time.

A real computer for every agent

Hardware-isolated microVMs boot in about 500ms with a shell, filesystem, and full browser. Pause, resume, and snapshots preserve state, while network controls keep your data secure.

Deploy agents to your customers

Ship agents under your own product, then expose them through the channels your customers already use — MCP servers, Telegram, APIs, and more. Deploy once, reach every surface.

Everything your agent needs to run reliably.

Production sandboxes with the primitives that make long-running agents practical — forking, streaming, and instant recovery out of the box.

Forking

Branch a running sandbox into parallel copies. Test variants, roll back, or fan out work — without reprovisioning.

Instant Boot

Cold start to running in under 3 seconds. No waiting for provisioning or image pulls.

~2scold start

Pause & Resume

Suspend sandboxes when idle — resume in milliseconds. Pay only for active compute.

resumed

0.4s

Browser Use

Give agents a full browser inside the sandbox. Navigate, click, fill forms, and extract data — headless or streamed live to your screen.

noVNC

live stream

Persistent Storage

Attached volumes survive restarts and snapshots. Your agent's data is always where it left off.

/data

/models

/cache

CLI & SDK Access

Full programmatic control. Execute commands, stream output, and manage sandboxes from code or terminal.

$ celesto computer run my-sbx "python train.py"

Run untrusted code and agents

Execute AI code and agents in disposable sandboxes—built for workflows, long-running agents, and computer-use agents.

$ pip install celesto $ celesto computer create --cpus 2 --memory 2048  Name:   einstein  Status: running $ celesto computer ssh einsteinroot@celesto:/# python3 -c "print('Hello from the cloud!')"Hello from the cloud! $ celesto computer run einstein "uname -a"Linux celesto 6.8.0 x86_64 GNU/Linux $ celesto computer delete einstein✓ Computer einstein deleted

No setup needed — sign up and start running

Works with every AI agent and framework

Celesto provides an open-source framework and CLI for lightweight, secure sandboxes to run AI-generated code.

LangSmith

OpenClaw

Hugging Face

Pydantic AI

OpenAI Agents

Docker

Qwen

Local LLM

Claude

NVIDIA

DeepSeek

Run a live service from the same sandbox.

Start Express, FastAPI, webhooks, dashboards, or an agent API inside an isolated computer. Publish a port and Celesto gives you a public URL - no separate deploy service to wire up.

No extra deploy service

Run the app and publish the port from Celesto. No Vercel-style detour or proxy glue.

A real public URL

Share an Express app, FastAPI server, dashboard, webhook receiver, or agent backend.

Still isolated

The service is reachable, but the runtime stays inside the disposable sandbox.

p-spark-8000.celesto.dev

Built and deployed with Celesto

A public web app, served straight from the sandbox.

Open live appnpmport 8000

$ npm run dev -- --port 8000

$ celesto computer port publish

port 8000

$ npm run dev -- --port 8000

ready on http://0.0.0.0:8000

$ celesto computer port publish

Port published

RunPublishShare

p-spark-8000.celesto.dev

Built and deployed with Celesto

A public web app, served straight from the sandbox.

Open live appnpmport 8000

Agent observability

Full observability, Know exactly why agents fail.

Every agent execution is automatically traced. Debug complex workflows, inspect tool calls, and optimize performance without writing a single log line.

Automatic Tracing

Built into Celesto: every agent call is traced automatically, no configuration needed.

Live Streaming

Watch traces stream in real-time as your agents run in production.

View observability dashboard

Example traceSuccess · 1.2s

input

action: get_order
ID: 12345

output

ID: 12345
status: shipped
carrier: FedEx
tracking: 7890123456

View raw JSON

input

{
  "action": "get_order",
  "id": "12345"
}

output

{
  "id": "12345",
  "status": "shipped",
  "carrier": "FedEx",
  "tracking": "7890123456"
}

Execution path

Shopify Tool

450ms

input

action: get_order
ID: 12345

output

ID: 12345
status: shipped
carrier: FedEx
tracking: 7890123456

View raw JSON

input

{
  "action": "get_order",
  "id": "12345"
}

output

{
  "id": "12345",
  "status": "shipped",
  "carrier": "FedEx",
  "tracking": "7890123456"
}

Usage-based sandbox pricing

Choose from six sandbox sizes and pay for running time. No monthly platform minimum.

Micro (default)

$0.015/hour

1 shared vCPU · 2 GB RAM · 20 GB root disk

Small

$0.020/hour

1 shared vCPU · 4 GB RAM · 25 GB root disk

Standard

$0.030/hour

2 shared vCPU · 4 GB RAM · 30 GB root disk

Memory

$0.040/hour

2 shared vCPU · 8 GB RAM · 40 GB root disk

Large

$0.080/hour

4 shared vCPU · 16 GB RAM · 60 GB root disk

XL

$0.160/hour

8 shared vCPU · 32 GB RAM · 100 GB root disk

Rates apply to new sandboxes; existing sandboxes retain their original rates. Prices are in USD; checkout shows your currency and any conversion fee. Start with a one-time $5 card-free trial for 30 days. Stopped root disks are retained free for 30 days.

Start free trial See full pricing details →

Give your agents aRealComputer

Computer using agents love Celesto

Your cloud computer, in three commands.

A browser your agent can use.

Step in. Then hand it back.

Pick up where your agent left off.

Open-source core. Self-host or use our cloud.

Celesto

Celesto Cloud

Build, run, and improve AI agents.

Agents that improve themselves

A real computer for every agent

Deploy agents to your customers

Everything your agent needs to run reliably.

Forking

Instant Boot

Pause & Resume

Browser Use

Persistent Storage

CLI & SDK Access

Run untrusted code and agents

Works with every AI agent and framework

Run a live service from the same sandbox.

No extra deploy service

A real public URL

Still isolated

Built and deployed with Celesto

Built and deployed with Celesto

Full observability, Know exactly why agents fail.

Automatic Tracing

Live Streaming

Customer Support Agent, agent1.2s

Router, system, child of Customer Support Agent0.1s

Shopify Tool, tool, child of Router450ms

input

output

Email Tool, tool, child of Router300ms

Execution path

Shopify Tool

input

output

Usage-based sandbox pricing

Micro (default)

Small

Standard

Memory

Large

XL

Give your agents aReal
Computer