kyora cloud

Kyora: vms, cloud and runtime for agents at scale.

Run coding agents, apps and containers at scale, each on its own machine, coordinated and observed, for less than any sandbox cloud we've compared.

try it live kyora cloud
01 / 02

vms

early access
kyora

a fresh vm · opencode inside · yours for a few minutes

pricing

billed per second · 2 tb bandwidth (in + out) free, plus 0.15 gb per vcpu-hour

$0.026an hour

one vm · 4 vcpu · 8 gb allocated · 50 gb nvme written

$0.024 compute + $0.002 disk

less per vm-hour than every sandbox provider we compared, without burst compare →

  • basea microvm in the nearest region
  • +25%pinned to a region
  • +50%best-effort burst use (capped by you)
  • +25%kyora ci: managed github actions runners
  • +50%ha / replicacoming soon
cpu
$0.0036per vcpu-hour allocated$2.628 per vcpu-month
memory
$0.00125per gb-hour allocated$0.912 per gb-month
disk
$0.025per gb-month of nvme written
bandwidth
$5per tb (decimal) in + out, after 2 tb free each month, plus 0.15 gb per vcpu-hour

inference

preview

from $2/h

priced by shares of gpus or whole clusters, not per token

set a cap like max $20/h, then run your gpus flat out

examples
  1. one vm, on all month4 vcpu · 8 gb allocated · 50 gb nvme written$19.06 a month
  2. a fleet: 1,000 subagents for 15 minutes2 vcpu · 4 gb allocated · 10 gb nvme written$3.14

effortless, highly performant and available compute, accessible in an agent native way

  • 01

    a vm for every agent

    • its own disk, network and cores
    • install anything, build, serve, drive a desktop
  • 02

    host anything

    • any port of a vm at its own url
    • on kyora.run or your domain, from cloudflare's edge
  • 03

    suspend · resume

    • memory and files in one snapshot
    • resume mid-flight on any machine in the region
  • 04

    detach · back

    • claude code or codex into a vm and back, one command
    • attach to watch it work; let go and it runs on
  • 05

    eu · us · apac

    • a named region is strict: eu stays in the eu
    • global: the nearest region with room
  • 06

    live screens

    • agents drive them with screenshots, mouse and keys
    • watch, take over, or play back a recording
explore more features
02 / 02

review

panel
  • codex
  • claude code
  • opencode
  • kimi
  • grok
  • qwen
  • glm
kyorabot · commentedexample

kyora · 4/5 · low risk

Retries partial writes in the export path and tightens its path checks.
Fix the retry offset before merging.

2 findings, both verified against the code

  • [major] Re-read the offset before retrying a partial write src/export/writer.rs:118consensus · 3 of 7
  • [minor] Keep the path in the error when the rename fails src/export/errors.rs:42verified · checked by qwen

Summary:

  • Retries partial writes up to three times, with backoff.
  • Rejects export paths that leave the workspace.

▸ Reviewer models

vision

the future of engineering, research and work at scale

A runtime orchestrating agent swarms at scale, each agent on its own vm. Dedicated inference capping cost to a minimum.

  1. 01

    the runtime

    rust core · rlm · scp

  2. 02

    vms

    scale · fork · sota perf

  3. 03

    inference

    preview
    gpugpugpu

    infinite capacity · 70-80% under api cost · pay for hardware, not tokens

01

research at scale

  • many agent runs at once, a vm each
  • fork a promising run from its snapshot, or reset it
  • open-weight models, unmetered coming soon