The missing runtime between an agent SDK and production

What it does

SandBase Harness is a local-first runtime that wraps model providers to give agents a production backbone. It stores agent definitions, session history, credentials, and memory in SQLite, exposes a Claude Managed Agents-style /v1 API with a local web console, and isolates tool execution inside local processes, Docker containers, or Kubernetes pods. Sessions can be replayed through resumable Server-Sent Events.

The interesting bit

The project refuses to be another model SDK or a visual workflow builder; instead it tackles the unglamorous infrastructure layer—governance, auditability, and sandbox boundaries—that turns a chat loop into a managed service. It also doubles as a native DeepSeek Harness plugin, bridging that ecosystem over stdio MCP.

Key highlights

• CMA-compatible /v1 API and local web console for inspecting runs.
• SQLite-backed metadata with pluggable sandbox backends: local process, Docker, Kubernetes, or self-hosted workers.
• MCP toolsets with permission policies, built-in tools, and skill packages.
• Resumable Server-Sent Events for session replay and debugging.
• Open-by-default auth that activates once API keys are configured.

Caveats

• The npm package name managed-agents is unscoped and currently belongs to a different project, so distribution is source-only for now.
• It requires Node.js 22+ and a model provider key, and the current release is v0.3.0, so expect a moving surface.

Verdict

Developers building autonomous agents that need audit trails, credential isolation, and session resumability should look here; if you just need a quick chat wrapper around an API, it is overkill.