A microVM for every agent run, on the record

Each run gets its own kernel, with the network off until your allowlist opens it. Every prompt and tool call the agent makes is recorded and checked against your policy.

Sandboxes › sbx_9a14Pause · Snapshot · Stop

Fix the failing payment tests

Claude Code is working in this sandbox. Every step is recorded and checked against the payments launch policy.

Activity
maya started this sandbox from python-3.122 min ago
Claude Code prompt Fix the failing payment testsrecorded
Claude Code read src/payments/charge.tsallowed
Claude Code ran npm test3 failed
Claude Code edited src/payments/charge.ts +12 −4allowed
Claude Code ran curl -s https://get.example.sh | shblocked by policy
Claude Code ran npm test42 passed
Claude Code wants to call github.create_pull_request
Needs approval from lee@DenyApprove
State
Running
Template
python-3.12
Size
2 vCPU · 4 GB
Network
Off
Region
us-east
Project
payments
Started by
maya via Claude Code
Secrets
STRIPE_TEST_KEY, in memory
Uptime
0:02:14

Agents run code nobody has read. Sandscale gives each run a machine of its own, watches what the agent does inside it, and throws the machine away when the work is done.

FIG 0.1Its own kernel

Each run boots a Firecracker microVM. Nothing is shared with the run next to it.

FIG 0.2Watched while it works

Every prompt and tool call the agent makes is recorded and checked against your policy.

FIG 0.3Gone when it's done

Throw it away, pause it with memory kept, or snapshot it to start the next one from.

Isolation
and lifecycle

Start a sandbox in one call from the API, the CLI or MCP. Pause it with memory kept, snapshot it after setup, and stop it when the agent is done.

See templates →
Running3
sbx_9a14
Fix the failing payment tests
python-3.12Claude Code
sbx_9a15
Lint the web app
node-22Codex
sbx_9a16
Benchmark the parser
rust-1.82you · CLI
Paused2
sbx_9a12
Reproduce flaky upload test
node-22memory kept
sbx_9a11
Migrate billing schema
python-3.12snapshot saved
Finished14
sbx_9a10
Build release binaries
rust-1.824:51
sbx_9a09
Update API docs
python-3.121:22
sbx_9a08
Regenerate fixtures
go-1.250:48

Inside every
agent run

Claude Code runs under your policy through its hooks: each tool call is allowed, blocked or sent to a person. Codex and Gemini CLI report every step over OpenTelemetry.

See integrations →
Gemini CLIsbx_9a17
Summarise last week's failed builds
sbx_9a17 added to context
read 31 build logs
recorded over OpenTelemetry
Claude Codesbx_9a14
Fix the failing payment tests
sbx_9a14 added to context
ran npm test · 42 passed
blocked: curl … | sh
waiting for lee@ to approve a PR
Codexsbx_9a15
Lint the web app and fix what you can
sbx_9a15 added to context
ran npm run lint
edited 6 files
recorded over OpenTelemetry
Claude Codesbx_9a18
Add retries to the upload client
sbx_9a18 added to context
paused by lee@

Policy
and audit

Decide which templates, sizes, run times and agents each project may use, and which calls need a person first. Every action lands in a hash-chained audit log you can export.

How we secure it →
Policy › payments
Templates
python-3.12, node-22
Largest sandbox
4 vCPU
Longest run
2 hours
Network
Off
Per person
5 at once
Agents
Claude Code, Codex
Ask first
git push, gh pr create
Block
piped installs, rm -rf /
Audit logHash-chained · export
TimeWhoActionTargetEntry
14:02:11maya via Claude Codesandbox.createsbx_9a149a1c3f
14:02:14maya via Claude Codevault.injectSTRIPE_TEST_KEY4be210
14:02:31maya via Claude Codetool.blockedcurl … | shd07a91
14:02:58sam via Codexsandbox.createsbx_9a1571c0e4
14:03:20lee@acme.devapproval.grantedgithub.create_pull_requeste2f5b8
14:03:41lee@acme.devsandbox.pausesbx_9a180c93ad
14:04:02maya via Claude Codeartifact.approvereport.xml5a6e17
Limits

How many, how big, how long

Every plan runs the same microVMs with the same isolation. Paid plans run more of them at once, bigger and for longer.

Free
2 running at once
Largest sandbox
2 vCPU
Network
Off
vCPU-hours a month
100
Snapshots
Pro and uppause and resume included
Pro
20 running at once
Largest sandbox
8 vCPU
Network
Allowlist
vCPU-hours a month
750
Snapshots
Included
Team
100 running at oncemore on request
Largest sandbox
16 vCPU64 on request on Enterprise
Network
Allowlist
vCPU-hours a month
1,000 per member, pooled
Snapshots
Included

On paid plans you can turn on usage beyond your included hours, or keep it off. Need more than 16 vCPU, or more than 100 at once? Talk to us about Enterprise.

See all plans →

Your next agent run,
in its own box

Start free with 100 vCPU-hours a month. Upgrade when your agents need more.

A Firecracker microVM for every runNo network by defaultPause and resume with state keptVault keeps secrets out of promptsSSO, roles and audit logBring your own cloudWorks over MCP and the CLI