Home
Why Sandlock
How It Works Use Cases Comparison Security Model
Docs
Documentation Home Getting Started CLI Reference Python SDK Sandbox Reference FAQ
Products
Overview Sandbox HTTP API Sandbox Scheduler
GitHub Schedule a Demo
Commercial offering

Operate Sandboxes at Fleet Scale, on Your Infrastructure

The open-source core confines one process. The Sandbox HTTP API and Sandbox Scheduler run thousands, fitting a thousand agent sessions into the memory a reservation model spends on twenty. Both deploy inside your trust boundary.

The split

Open Source at the Core, Commercial at the Fleet Layer

Everything that confines a process is Apache-2.0. Everything that operates a fleet is licensed. No paywalled flags, nothing that phones home.

Sandlock

Apache-2.0

The sandbox, the supervisor, copy-on-write, HTTP ACLs, credential injection, handlers, and every interface: CLI, SDKs, OCI runtime, MCP server. Free at any scale.

Who this is for

Teams Running Sandboxes as a Fleet

One sandbox on one host? The open-source project is the whole answer. This layer starts when sandboxes outnumber machines.

AI agent platforms

A sandbox per session, and a memory bill dominated by sessions doing nothing.

Code execution products

Notebooks, interpreters, and eval harnesses that need tenancy, quotas, and per-key accounting.

Developer platforms

CI and build services where per-job VMs are the line item you want gone.

Security engineering

Approving untrusted execution on shared hosts, with evidence to hand an auditor.

HTTP API

A Sandbox Service Running on Your Own Servers

The core is a function call. The API makes a sandbox a resource: an ID, a lifetime, any language, and keys with scopes.

Any language, scoped keys

REST, so callers need no Linux and no FFI. Keys carry scopes and per-key usage accounting.

Sandboxes that outlive the call

Stable IDs, reconnect from another machine, named checkpoints to stop and restore.

Container images as the rootfs

Name an image; the server extracts, caches, chroots, and applies copy-on-write.

Long-running services

Start a server in a sandbox, query its ports, tail its output, stop it.

Remote MCP

Point Claude Desktop, Cursor, or any MCP client at the API. Each tool call gets its own sandbox.

Evidence for your auditor

Posture reports and tamper-evident audit trails, generated on your infrastructure.

Sandbox Scheduler

Provision for Execution, Not for Idle Sessions

An agent session spends nearly all of its life waiting on the model. The Scheduler keeps it in a checkpoint store and gives it node memory only while it runs.

1,000

Sessions

At 2 GB each, reservation holds 2 TB whether or not anything runs.

42 GB

Actual footprint

About ten execute at any instant. Provision for p99.9 concurrency and the same thousand fit.

~48×

More sessions per host

Approaching 91× at a hundred thousand sessions, as concurrency averages out.

Modelled from the duty cycle. We size the deployment against your own session count, concurrency, and idle profile.

The session lives in the store. Nodes checkpoint it when it goes quiet; any node with room restores it sub-second.
Node failure is not an incident. Failover, eviction, and rebalancing are the same path: restore elsewhere.
Forking is free. Restore one checkpoint under many IDs, which is what best-of-N and tree search need.
Two things do not move. TCP connections drop on relocation and GPU state does not checkpoint.
Deployment model

Your Infrastructure, Not a Shared Cloud

Licensed software, not a hosted service. Your workloads never cross a vendor boundary.

Your account or your data center

AWS, GCP, Azure, or bare metal. Nothing needs KVM, so ordinary cloud VMs are fine.

No dependency on us

No hosted control plane, no license server, no telemetry. Our outage cannot become yours.

An open-source exit

Drop the commercial layers and the complete Apache-2.0 core remains yours.

Before you book

The Questions Every Evaluation Raises

Sandlock shares the host kernel. Is that acceptable for untrusted code?

It is a kernel-enforced allowlist, not a VM. If your threat model includes a kernel exploit, run Sandlock inside the VM boundary you already trust; the Scheduler packs density inside it.

Why not Firecracker, or a hosted sandbox cloud?

Firecracker needs KVM and holds memory for every idle microVM. Hosted clouds run your customers' code on someone else's kernel. Sandlock runs unprivileged in your account and assigns memory only while a session executes.

Our fleet is not on Linux 6.12 yet.

The strict default refuses to start when a protection is missing. Each one can be waived explicitly and shows up in the posture evidence. Bring the kernel mix to the call.

What happens if we stop paying?

The API and Scheduler stop being licensed. The sandbox, SDKs, CLI, OCI runtime, and MCP server are Apache-2.0 and keep working. There is no license server to cut you off.

Commercial terms

What Is Free, What Is Licensed, and How It Is Priced

Sandlock

Apache-2.0
Free
Includes
Core, CLI, Python, Rust, and Go SDKs, OCI runtime, MCP server.
Limits
None. No caps, no telemetry, no registration.
  • Self-hosted only. No hosted or shared-tenant offering.
  • Pilot first. On your infrastructure, against real traffic, before anything is signed.
  • Support included. Engineering support and a response SLA with every license.

Evaluate It Against Your Workload

A thirty-minute engineering call with the CEO, sized against your numbers, then a pilot on your infrastructure. Bring the workload.