Fast, Multi-Cloud Sandbox Engine for AI Agents
Agent Sandbox is an open-source sandbox engine for AI agents. It runs on Kubernetes, serves the E2B API, and gives an agent, an evaluation or a training loop its own isolated environment in the time it takes to swap a container image.
Sandboxes are claimed, not created. A pool holds Pods that are already running, so serving a request is an in-place image swap rather than a scheduling round-trip. Capacity is a number you decide — a pool's replica count — and the autoscaler moves it as demand changes.
There is no build step. An E2B template is a snapshot compiled from a Dockerfile; here you name a container image when you create the sandbox, and the platform swaps it in. Any image your cluster can pull is a workload.
Platform and sandboxes are two surfaces. abx addresses the platform —
environments, warm pools, autoscaling, quotas, templates — and the E2B SDK
addresses the sandboxes themselves. The line is deliberate: sandbox work needs
no new client, and platform work needs no SDK.
helm upgrade --install agent-sandbox-worker \
oci://ghcr.io/scitix/agent-sandbox-worker \
--namespace agentbox-system --create-namespace \
--set controller.localClusterId=YOUR_CLUSTER_IDThat release is the whole platform on one cluster: the CRDs, the controller and
API, and the data plane sandboxes are reached through. localClusterId is
required — it is how a pool is placed on a cluster segment.
The console is a second, optional release that reads the same API:
helm upgrade --install agent-sandbox-hub \
oci://ghcr.io/scitix/agent-sandbox-hub \
--namespace agentbox-system \
--set env.secret="$(openssl rand -hex 32)" \
--set clusters[0].id=YOUR_CLUSTER_ID \
--set clusters[0].url=http://agent-sandbox-api.agentbox-system.svc.cluster.local \
--set ingress.host=agentbox.example.com→ Installation guide — prerequisites, the services you get, the shared secret between the two releases, and how to remove it all again.
curl -fsSL https://oss-ap-southeast.scitix.ai/scitix/packages/agentbox/cli/latest/install.sh | sh
abx context set YOUR_DEPLOYMENT \
--endpoint 'https://YOUR_CONSOLE/agentbox' \
--api-key agbx_...
abx clustersabx is a single binary with no runtime dependencies. The same script writes
nine skills to ~/.agents/skills — the vocabulary an agent reads before it
touches the platform, from abx-common (endpoint, key, cluster, approvals) to
abx-sandbox-secrets. Claude Code users can install the plugin instead:
/plugin marketplace add scitix/agent-sandbox.
→ CLI guide · Skills · E2B SDK guide
Every page is also a plain-text document: append .md to any docs URL
(/docs/concepts/pools.md), or hand an agent
llms-full.txt.
RL training requires thousands of environment resets per hour. Agent Sandbox pre-warms a pool of sandboxes so each rollout worker gets a fresh, isolated environment in milliseconds — removing the environment-reset bottleneck from your training loop. Supports SWE-bench Verified, SWE-Gym, Terminal-bench, and custom task distributions.
Give every agent turn or eval call its own isolated execution environment. The E2B-compatible API means existing SWE-agent, SWE-ReX, and similar frameworks work without modification.
Deploy sandbox pools across multiple clouds or regions. The built-in ExtProc routing layer dispatches requests to the most available cluster transparently — no routing logic required in application code. Supported cloud providers: AWS, Google Cloud, Azure, Alibaba Cloud, Volcengine, Cloudflare.
Coming soon: microVM-backed sandboxes for stronger isolation guarantees.
| Resource | Link |
|---|---|
| Documentation site | scitix.github.io/Agent-Sandbox |
| Installation — Helm, the CLI and the skills | /docs/installation |
| Concepts — the object model, envs, pools, autoscaling | /docs/concepts |
CLI guide — abx, endpoints and keys |
/docs/tutorials/cli |
| E2B SDK guide — creating and driving sandboxes | /docs/tutorials/e2b |
| Skills — the nine files an agent installs | /docs/skills |
| Examples — templates and environments to apply | /docs/examples |
| Designs — the egress filter, cross-cluster routing, our envd patches | /docs/designs |
| Managed agents — the Brain/Hands split | /docs/tutorials/managed-agents/brain-hands |
| API Reference (OpenAPI) | /docs/api/sandboxes/CreateSandbox |
Contributions are welcome — bug reports, feature requests, documentation, and code. Please read CONTRIBUTING.md before submitting a pull request.
All commits must include a Signed-off-by line (DCO). Use git commit -s.
Apache License 2.0 — see LICENSE for details.
Copyright © 2026 ScitiX.