Our take
- Activity
- 90
- Community
- 83
- Issues
- 85
- Pull requests
- 62
+1,543 stars in the last 30 days
- Releases more often than 93% of the projects we track
- Merges more pull requests than 90% of the projects we track
- Closes issues faster than 81% of the projects we track
Project history
Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.
Stars
17,496Forks
1,687Change since Nov 17, 2026: +105 stars·+21 forks
Overview
OpenSandbox is a runtime and API for giving AI applications isolated environments to execute code, run commands, manage files, and operate browsers or desktops. It addresses the need to run agent-generated actions away from the host application while managing sandbox creation and cleanup through a consistent interface.
It is intended for teams building AI agents and agent infrastructure who want to host these environments themselves. Local Docker use supports development and evaluation, while Kubernetes deployment targets managed or larger-scale workloads.
Key Features
- Create and manage sandboxes through a shared lifecycle API, with Docker locally and Kubernetes as a deployment option.
- Execute commands and scripts, and read or write files through built-in APIs and SDKs.
- Integrate from Python, Java/Kotlin, JavaScript/TypeScript, C#/.NET, or Go.
- Apply inbound routing and per-sandbox outbound network policies.
- Use a Credential Vault to let agents access external services without exposing real credentials to sandbox workloads.
- Connect through a command-line interface or an MCP server.
- Support browser and desktop automation, coding-agent workflows, and evaluation workloads through examples and integrations.
Use Cases
- Agent platform teams can provide isolated working environments for agents that edit files and run commands.
- Developers can test code-execution or data-analysis agents without running generated code directly in the host application.
- Teams building browser or desktop agents can run automation in sandboxed environments.
- Evaluation and reinforcement-learning teams can create separate sandboxes for trials and batch workloads.
Getting Started
Requirements for the local setup are Docker and Python 3.10 or later. Initialize a Docker example configuration and start the server:
uvx opensandbox-server init-config ~/.sandbox.toml --example docker
uvx opensandbox-server
Install the Python SDK with uv pip install opensandbox. See the README for the complete example, configuration, and deployment guidance.
Alternatives
- SandBase Harness: SandBase Harness focuses on local-first agent sessions with memory, credentials, and auditing, rather than a shared Docker and Kubernetes sandbox API.
- Wassette: Wassette isolates WebAssembly components exposed through MCP, rather than providing general-purpose code, command, browser, and desktop sandboxes.
| Project | Language | License | Stars | Status |
|---|---|---|---|---|
| OpenSandbox | Python | Apache-2.0 | 15.7k | Not checked yet |
| sandbase-harness | TypeScript | Apache-2.0 | 640 | New |
| wassette | Rust | MIT | 972 | Active |
Considerations
- Local execution requires Docker. The README also lists Python 3.10 or later for examples and the local runtime.
- Kubernetes deployment involves additional operational setup. Consult the project deployment and architecture documentation before planning a cluster rollout.
- This is infrastructure for running agent workloads, not an agent or model by itself. Integrators are responsible for configuring access, network policies, credentials, and deployment security for their environment.
- The repository is active and substantial, but the supplied metadata reports 140 open issues. Review current issues and project guidance when assessing fit.
Found this useful?
Share it with someone who would like OpenSandbox.