Kanvify boots isolated microVMs in seconds — persistent machines your AI agents live in, and a fresh one for every CI job. Governed from outside the box. Billed by the second.
Plays well with the tools you already run
Your agents need a computer that keeps its state. Your pipeline needs one that leaves nothing behind. Kanvify runs both on the same metal, the same isolation model and the same meter.
Persistent sandboxes that keep their files between sessions. Suspend when idle, resume exactly where the agent stopped, and checkpoint before anything risky.
Ephemeral microVMs pick up your GitHub Actions jobs, run real Docker, and are destroyed at exit. There is no standing fleet to secure.
Same metalSame isolationOne meter
$0.10 / vCPU-hour for bothOne API call boots a machine. The lifecycle takes it from there — and compute stops whenever your agent does.
Controls sit between the agent and the world — not inside the prompt, where a clever agent could talk its way around them.
Per-sandbox allow and deny rules. If a domain isn’t on the list, the connection never leaves the box.
Model calls route through the proxy with your provider key — budgets enforced per call, caching and fallbacks included, no Kanvify LLM charge.
A point-in-time image of the disk. Roll back to a known-good state when a run goes sideways.
Metered per second, attributed per sandbox, and stopped by an account-level cap before runaway usage becomes a bill.
Sandbox activity is logged to an audit trail you can review after the fact.
Everything on this page is an API call.
# one static binary, standard library only $ export KANVIFY_API_KEY=knv_… KANVIFY_WORKSPACE=ws_… $ kanvify create --image ubuntu:22.04 --name agent --cpu 2 --memory 4 sbx_Q7mK2vP9xLr4 agent running $ kanvify cp ./task.md agent:/workspace/task.md uploaded ./task.md -> agent:/workspace/task.md $ kanvify exec -s agent -- make test # exits with the real code $ kanvify exec -s agent --interactive # or just drop in
$ curl -X POST "$KANVIFY_BASE_URL/workspaces/$WORKSPACE_ID/sandboxes" \ -H "Authorization: Bearer $KANVIFY_API_KEY" \ -H "Content-Type: application/json" \ -d '{"sandbox": {"name": "agent", "image": "ubuntu:22.04", "cpu": 2, "memory_gb": 4}}' { "sandbox": { "id": "sbx_Q7mK2vP9xLr4", "status": "creating", … } }
Managed GitHub Actions runners on ephemeral microVMs — a real machine per job, real Docker, and a spec sheet instead of a slogan. We publish our numbers, including the ones we’re still improving.
One command from your Kanvify workspace connects your GitHub org. Nothing to host, patch or scale on your side.
Change runs-on: to kanvify in any workflow. Your steps, actions and secrets stay untouched.
Each queued job boots a microVM that registers just-in-time. Cold queue-to-pickup is about 60 seconds today; we publish it because we intend to shrink it.
jobs: build:- runs-on: ubuntu-latest+ runs-on: kanvify steps: - uses: actions/checkout@v4 - run: docker compose up -d # real dockerd — just works - run: pnpm test
A 61-second job costs 61 seconds. No rounding up, no included-minute buckets.
Rounds each job up to a full minute · shared-pool VM
Mostly per-minute · varying isolation
Per-second — fresh microVM per job, destroyed at exit
Per-minute · ephemeral VM, more DIY
Comparison for a 2 vCPU x64 Linux runner. Market rates as commonly published, 2026-07.
Nothing is reused; nothing is left behind for the next run to find.
GitHub-hosted at $0.006/min vs. Kanvify at $0.0033/min — 2 vCPU × $0.10/vCPU-hour, billed per second.
You save
$26.67/ month
$320.00/ year
New accounts start with $5 of compute credit (about 1,500 runner minutes, valid 90 days). Card required.
Estimates cover 2 vCPU runner compute only. They exclude storage and cache (billed separately at $0.0001/GB-hour) and GitHub’s included free minutes. GitHub-hosted rate as published January 2026.
Compute, sandbox storage, and cache storage are three separately metered invoice lines. Bring your own LLM provider key; Kanvify does not charge for model usage.
Meter 01 · Compute
$0.10/ vCPU-hour
vCPU-hours while a sandbox is running. Ephemeral CI runners and persistent agent computers use the same rate; suspended computers do not accrue compute charges.
Meter 02 · Sandbox storage
$0.0001/ GB-hour
GB-hours for sandbox filesystems and retained checkpoints. Retained checkpoints continue to accrue storage usage after their source sandbox is suspended or destroyed.
Meter 03 · Cache storage
$0.0001/ GB-hour
GB-hours for retained compressed CI cache data. Cache uploads, downloads, transfer bytes, and R2 operation counts are not billed.
LLM proxy · Bring your own key
You pay your provider directly. The proxy still adds caching, routing, cost visibility, and keeps your key out of the sandbox.
No tiers. No seats. No minimum. Pay for what you run.
We list what we don’t do so you can trust what we say we do. These are on the roadmap; nothing above is.
Boot a sandbox, hand it to your agent, point a workflow at runs-on: kanvify. Self-serve, and billed only for what you run.
Self-serve. Spend caps on by default.