sandbaseai/sandbase-harness

What it solves

SandBase Harness is a local-first runtime for AI agents that provides the necessary infrastructure for production-grade agents. While most SDKs only handle the model loop, Harness provides the operational layer including persistent sessions, sandboxed code execution, credential management, and audit trails, allowing developers to run agents on their own machine or infrastructure without requiring a hosted control plane.

How it works

It operates as a runtime layer that manages the lifecycle of agents and their environments. It uses a SQLite-backed system to store metadata for agents, sessions, and memory, while providing a /v1 API (compatible with Claude Managed Agents) and a built-in Console for human inspection. The system supports multiple sandbox backends for safe code execution, including local processes, Docker containers, and Kubernetes, and integrates with the Model Context Protocol (MCP) to provide toolsets and skills.

Who it’s for

AI agent developers who need a secure, auditable, and self-hosted environment to run their agents, particularly those who want to avoid reliance on a hosted control plane and need fine-grained control over tool access and code execution sandboxes.

Highlights

• Local-First Infrastructure: Uses SQLite and local file storage for metadata and state, keeping data on the user's machine.
• Flexible Sandboxing: Supports multiple execution environments including Docker (per-session containers) and Kubernetes.
• Multi-Model Support: Compatible with OpenAI, Anthropic, MiniMax, and OpenAI-compatible providers like DeepSeek V4.
• Operational Tooling: Includes a built-in Console for session replay, debugging, and auditing agent actions.
• MCP Integration: Built-in support for MCP toolsets and a bridge for integrating with other MCP clients.