Backed by Y Combinator

Give every agent an ultrafast computer in the cloud. Boot in seconds, pause and resume for free, and connect GPUs. Starting from $0.0306 per hour.

curl -fsSL https://www.beam.cloud/install | sh -s -- --vm

Running in US East · 4 vCPU · 8 GiB · 50 GiB disk$0.144/hr · $0 while stopped

Powering products and agents at

Performance

Agents, Meet Serverless

Beam is building the first bare-metal cloud for agents. We provide persistent, forkable microVMs you can expose, pause, network-isolate, and GPU-accelerate, billed only while they run.

Resumed in ~1s

RUNNING

PAUSED

RESUMED · ~1s

RAM12.4 / 16 GiB · restored

PIDPROCESSMEMSTATE

412python agent.py6.1Grunning

977node server.js1.8Grunning

1203postgres2.9Grunning

1388chromium --headless1.6Grunning

Memory snapshots

Stateful pause and resume

RAM and running processes resume in under a second.

Disk usage (GiB)

Used 23.4 GiBLimit 100 GiBBilled 23.4 GiBPERSISTENT

data/11.0G

.cache/hf/6.0G

.venv/5.0G

repo/1.4G

Usage-based storage

Fast, persistent local disks

Up to 100Gi per machine. Only pay for what you use.

Clock speed (GHz)

Cores 16Threads 32Boost 5.7 GHz

Your VM · bare metalup to 5.7 GHz

Typical cloud vCPU~2.6–3.6 GHz

Raw performance

High-performance CPUs and GPUs

S-tier hardware, virtualized on bare metal.

agent-7f3a.app.beam.cloud

Requests 21Awake 15 minAsleep 45 min

BILLED COMPUTE · LAST HOUR

Scale to zero4× less15 min

Always on60 min

Pay for what runs

Scale to zero

Idle VMs pause and wake automatically when invoked.

Platform

The only cloud your agent needs

Agents drive every VM over MCP, while keys, network rules and URLs stay under your control. GPUs, images and volumes are on the same platform.

MCP connected

MCP tools 100+Calls 0/6

create_vmubuntu-24.04 · 4 vCPU…

fork_vmvm-7f3a → 3 forks

run_commandpytest -q

screenshotdesktop · 1280×800

clickx 412 · y 268

type_text"ship it"

MCP server

The entire cloud, at your fingertips

100+ MCP tools to create, fork, run commands, and drive the desktop: screenshot, click, type.

Outbound request

vm-7f3ainside the VM

$ echo $OPENAI_API_KEYplaceholder_openai

Beam egresskey swap

✓api.openai.comreal key added

✕any other hostplaceholder only

responseskey removed

workspace token in VMnone

Egress proxy

Secrets never enter the machine

Your keys are proxied to the VM, so real secrets are never exposed in the machine.

egress.rules

microVMHARDWARE ISOLATED

RULEHOSTHITS

allowapi.openai.com12

allowpypi.org48

allowgithub.com3

deny*7

LAST BLOCKED

✕169.254.169.254metadata

✕paste.example.comdenied

Hostname egress rules

Hardware isolation

Lock a box to just the APIs it needs.

vm-7f3a · network

BEAM

2001:db8:7f3a::1csame address

ELSEWHERE

203.0.113.8198.51.100.23203.0.113.91

a new address after every resume

AAAAagent.example.comcustom domain

ALLOW2001:db8:7f3a::1c/128allowlist

IPv6

Persistent IP address per VM

Use it for custom domains and allowlisting.

Protected

https://agent-7f3a.app.beam.cloud

GET /no token401

GET /?share=…share link200

GET /?share=…after 1 hour401

Share link1 hour

…app.beam.cloud/?share=k9x2f7

expires in58:41

Share links

Batteries-Included Auth

Every URL is private until you share it, and share links last an hour.

Images & storage

ghcr.io/acme/agent:latest✓

SNAPSHOTS

restoresnap-4→vm-4

Images, volumes, snapshots

Bring your own image

Boot from your own image, share volumes across VMs, and snapshot any of them.

GPU catalog

GPU types 29

ServerlessScale to zero

ReservedDedicated machines

Serverless and reserved

Thousands of GPUs

Over 30 GPU types available on-demand.

Everything You Need

Durable disksDatabases, Redis and NVMe drives.

StacksA visual canvas for your infrastructure.

Inference endpointsHigh-performance endpoints for GPU inference.

Storage volumesGlobally distributed storage volumes.

Autoscaling

Hey Codex, scale to users

Autoscale to millions of concurrent users without setting up your own servers or managing a Kubernetes cluster. Your agent shipped, so let us cook.

Examples

What will you build?

Make your vision come to life. Beam manages the VMs, sandboxes, and GPUs powering your app, so there's zero infrastructure to manage.

All use cases

devbox · terminal

beam vm new devbox
beam vm sync devbox ./app
beam vm exec devbox -- npm test

# Share what the agent built
beam vm expose devbox 3000 --protected
beam vm open devbox 3000

Coding agents

Give every coding agent a dev box

A persistent Linux VM with SSH and a durable disk, so repos and tools persist between sessions.

desktop-dev · desktop

from beam import VM

with VM.connect("desktop-dev") as vm:
    vm.desktop.screenshot("before.png")
    vm.desktop.click(200, 150)
    vm.desktop.press("ctrl+l")
    vm.desktop.write("news.ycombinator.com")
    vm.desktop.press("Return")

Computer use

Drive a real desktop

Agents screenshot, click, and type in a full Linux desktop through the SDK or MCP.

agent.py

from beam import Image, Sandbox

image = Image(python_packages=["pandas"])
sandbox = Sandbox(image=image).create()

result = sandbox.process.run_code(agent_code)
print(result.result)

sandbox.terminate()

Code execution

Run agent-generated code safely

Untrusted code runs in an isolated sandbox that boots in seconds and shuts down when it's done.

rollouts.py

snap = base.create_image_from_filesystem()
image = Image.from_id(snap)

envs = [Sandbox(image=image).create()
        for _ in range(16)]

for env in envs:
    env.process.exec("python3", "rollout.py")

RL environments

Fan out rollouts from one snapshot

Every episode starts from the same saved state, in parallel, with GPU training on the same platform.

agent · lifecycle

beam vm new agent
beam vm update agent --ttl 300 \
  --idle-action pause --auto-resume
beam vm expose agent 8000 --protected

# 5 idle minutes: paused, RAM saved
# Next request: resumed in about a second

Agent hosting

Host agents that sleep when idle

The VM pauses after five idle minutes with its memory intact, then wakes when its URL is hit.

app.py

from beam.integrations import VLLM, VLLMArgs

llm = VLLM(
    name="qwen-agent",
    gpu="H100",
    vllm_args=VLLMArgs(model="Qwen/Qwen3-8B"),
)

# beam deploy app.py:llm

Inference

Serve models next to your agents

Deploy open models with vLLM on GPUs behind an OpenAI-compatible URL that scales to zero.

Community

What developers say

Join thousands of developers who moved production workloads to Beam and stopped managing their own infra.

Beam has been a huge time-saver by eliminating the need to monitor and manage my own VM infrastructure.

I no longer worry about unexpected bugs or outages which means less downtime and fewer headaches.

This lets me provide a significantly more reliable service to my users, and it's been surprisingly more cost-efficient than my prior solution.

Liam EloieMachine Learning Engineer

One of the better developer experiences I've had in a while was with @beam_cloud - a serverless GPU and API infra platform. Check them out 👇

Deploy an open source model on hugging face running on GPUs in a few minutes with 6 lines of code.

Keep your eyes on these guys 👀

Brandon Garcia@__BCG__

Beam has been a revelation in terms of making it simple to build an ML application on GPU

Devon PeroutkySoftware Engineer

Frase is running language models exclusively on Beam and it was surprisingly easy to migrate, less maintenance, and is saving us money because unlike Google and other cloud providers, Beam is able to provide us with an on-demand solution that scales immediately with our traffic, and we don’t need to worry about any of the clunky tooling around GPUs.

Frankie L.CTO and AI Researcher @ Frase

@beam_cloud is 🔥. Such a huge workflow improvement over AWS Sagemaker / Google vertex ai

Eric Meier@bitphinix

Spun up a new app today and realized just how it easy it was. Took me only 15 mins to organize and deploy on Beam.

Realizing that quick python apps on Beam is a cheat code

Brandon BrisbonCTO at Shop Galaxy

Beam is amazing. I tested the CLI and in 5 minutes had something running on the cloud.

And the Slack community is a game changer because when we get stuck we get responses quickly

Leonardo CucoCTO at Ween.ai

Beam is powering hands-down the best developer experience to run models on GPUs easily at scale. Best decision on the infra side for us this year so far.

Louis MorgnerCo-founder, AI lead @ Jamie

Time is the biggest thing Beam has helped us with. I went from spending 6 hours developing an API to pressing a button and deploying instantly

Benjamin SmithMLE at Shippabo

I can't recommend Beam highly enough. Their developer experience is top notch.

We never could have shipped Happy Accidents as quickly as we did without them. We were able to build the GPU portion of our app in hours instead of weeks.

Not only is the platform great, we loved working with the Beam team. They're extremely responsive, so we had a high level of confidence in the reliability of the platform.

James BonnerFounder at Happy Accidents

If you're looking to dip your toes into building something with AI, definitely take a look at http://beam.cloud.

Serverless functions with access to GPUs so you can run jobs on-demand and pay only for what you use.

And it's *much* easier than setting up a VM somewhere!

Joshua Clanton@joshuacc

Pay as you gobilled by the millisecond

Give Your Agent Superpowers

Start shipping today and bring your vision to life.

Start Building Read the docs

High-Performance Compute for Agents