SandBase Harness
SandBase Harness is an open-source, local-first managed agent runtime designed for running AI agents with persistent sessions and sandboxed execution. It is CMA-compatible and works with any model, providing a unified layer for agent orchestration. The runtime includes built-in support for MCP (Model Context Protocol) tools, enabling agents to access external services and data sources through a standardized interface. A key feature is its human console, which allows operators to monitor, audit, and replay agent sessions. This provides transparency and control, essential for debugging and compliance. Sandboxed execution ensures that agent actions do not affect the host system, reducing risk when running untrusted or experimental code. Additionally, SandBase Harness ships with a native DeepSeek Harness bundle over stdio MCP, making it straightforward to connect DeepSeek models as agent backends. The project is available on GitHub and is designed for developers and teams who want to run self-hosted AI agents with robust safety and observability features.
- Category
- AI Agents
- Pricing
- Open Source
Problem solved
SandBase Harness addresses the need for a safe, auditable, and model-agnostic runtime for AI agents. It provides persistent sessions, sandboxed execution, and MCP tool integration, along with audit and replay capabilities. This solves the challenge of deploying agents that can interact with tools and external systems while maintaining oversight and reproducibility, all within a local-first environment.
Use cases
- Run local AI agents with persistent memory and sandboxed tool execution
- Integrate MCP-compatible tools for agent workflows
- Audit and replay agent sessions for debugging and compliance
- Provide human oversight through a local console
- Deploy DeepSeek models as agent harnesses over stdio MCP
- Develop CMA-compatible agents with any model backend
Audience
- Developers
- ML Engineers
- Data Scientists
- Researchers
Pros
- Local-first design ensures data privacy and low latency
- Sandboxed execution isolates agent actions from the host system
- Built-in audit and replay enable debugging and reproducibility
- MCP tool support allows integration with a wide range of external services
- Model-agnostic and CMA-compatible, supporting any model including DeepSeek
Cons
- Requires self-hosting and local infrastructure management
- May have a learning curve for configuring MCP tools and agent sessions
- No managed cloud version mentioned, limiting non-technical users
Official website: https://github.com/sandbaseai/sandbase-harness
Canonical LambdaBase page: https://www.lambdabase.com/tools/sandbase-harness