AI Native Landscape

SandBase Harness

Tracked

A local-first, self-hosted runtime for building and operating AI agents, providing persistent sessions, sandboxed execution backends, credential vaults, approval gates, audit trails, and replay.

Author SandBase AI Open Sourced 2026-07-11 Last Commit Unknown

Overview

SandBase Harness is a local-first runtime layer that sits between agent SDKs and production agent deployments. Agent SDKs handle the model loop; Harness adds what production agents actually need — persistent sessions, tool governance, sandbox boundaries, credential handling, memory, auditability, and a local Console UI for humans to inspect what happened. It runs on a machine or in self-owned infrastructure with SQLite and file storage and no required hosted control plane.

Key Features

  • Claude Managed Agents-style /v1 HTTP API plus a built-in local Console dashboard
  • Persistent, resumable sessions exposed as Server-Sent Events for replay and debugging
  • Sandbox execution backends: local process, Docker (per-session containers), Kubernetes (kubectl exec/cp), and self-hosted worker queue
  • MCP toolsets, permission policies, approval gates, built-in tools, and skill packages
  • SQLite-backed agents, sessions, environments, credential vaults, memory stores, files, and API keys; file/skill bytes stored in the workspace state directory
  • One active model provider boundary (OpenAI, Anthropic, MiniMax, or OpenAI-compatible endpoints) configured through Settings V2 with validation and restart flow

Use Cases

  • Running generated code safely with a selectable isolation boundary from laptop (local process) to cluster (Kubernetes)
  • Inspecting long-running agents: resuming event streams, auditing tool calls, and replaying what an agent did
  • Controlling tool access for third-party plugins with credential vaults, permission policies, and human approval gates
  • Operating an auditable coding or research agent entirely on self-hosted infrastructure

Technical Details

  • Node.js 22+ / TypeScript codebase; TypeScript SDK at managed-agents/sdk and a release gate via npm run release:check
  • Settings V2 pins one workspace model vendor, loop engine, storage, memory, and sandbox backend, with form/JSON editing modes and a validation-driven restart flow
  • The DeepSeek Harness bridge runs over MCP stdio: installing the bundle into a DSH profile exposes native mcp__sandbase__* tools for listing agents, creating and running sessions, inspecting results and artifacts, and stopping work
  • Deployment as a DSH plugin resolves the tagged HTTPS Git source directly (running prepare only when dist/ is missing), with documented pnpm allowBuilds and duplicate-install recovery paths
  • State model separates SQLite metadata from file/skill bytes in the workspace state directory, keeping local-first storage without a hosted control plane
  • Distributed as a source checkout (v0.3.8 tag), an MCP stdio image, and an official MCP Registry entry