sandbase-harness: Run Self-Hosted AI Agents with Sandboxed Tools

Overview

SandBase Harness is a local-first, self-hosted runtime for AI agents, with a built-in Console and an MCP bridge. It addresses the operational needs around agent SDKs: persistent sessions, sandboxed tool execution, memory, credential handling, and audit trails. Its /v1 API follows Anthropic's Claude Managed Agents resource shapes, while the runtime supports multiple model providers rather than depending on a single model.

It is aimed at developers and operators who want to run and inspect agents on their own machine or infrastructure instead of relying on a hosted control plane. SQLite and file storage are the defaults, with sandbox backends for local processes, Docker, Kubernetes, and self-hosted workers.

Key Features

• Persistent agent sessions with resumable Server-Sent Events, audit logs, and replay.
• Sandboxed execution through local, per-session Docker, Kubernetes, or self-hosted worker backends.
• MCP toolsets, permission policies, approval gates, and credential vaults.
• Memory stores, files, skills, and session resources managed through the runtime.
• Local Console and a /v1 API compatible with the Anthropic TypeScript SDK's client pattern.
• TypeScript SDK and CLI for managing agents, environments, deployments, and chat sessions.
• SQLite metadata and local workspace storage by default, with API-key authentication available.

Use Cases

• A developer building an agent that needs persistent sessions and human approval for sensitive tool calls.
• A team operating coding agents that execute generated code in Docker or Kubernetes sandboxes.
• An operator who needs to review, replay, and audit agent activity on infrastructure they control.
• A DeepSeek Harness or other MCP client user who wants to expose the runtime's tools over an MCP bridge.

Getting Started

Requires Node.js 22 or later and npm 10 or later. From the repository, build the runtime and initialize a workspace:

git clone --branch v0.3.8 --depth 1 https://github.com/sandbaseai/sandbase-harness.git
cd sandbase-harness
npm ci && npm run build
mkdir ../my-agents && cd ../my-agents
node ../sandbase-harness/dist/index.js init
node ../sandbase-harness/dist/index.js start

Configure a model provider and model ID in the Console at http://127.0.0.1:3000/dashboard. See the installation guide and README for setup details.

Alternatives

• maka: Maka is a macOS-focused agent workspace with an append-only runtime log, rather than a general local-first runtime and MCP bridge.
• swarmclaw: SwarmClaw focuses on coordinating agent teams with schedules, delegation, and external runtimes, rather than providing a managed-agents-style API and replay.
• OpenBot: OpenBot centers on AI coworkers with browser and file access through AG-UI, rather than a local-first agent runtime with an MCP bridge.

Considerations

The project is new and described by its health data as changing fast. It has recent commits and releases, six contributors, and is not marked as dependent mostly on one person, but its short history still leaves less time for operational behavior to be established.

Docker is optional for the basic setup, and there are no GPU or Compose requirements listed. Sandboxing and network controls vary by backend: the README says limited networking is unsupported on Kubernetes and self-hosted workers, and local-process proxy enforcement is best-effort. The project also explicitly does not implement threads or coordinator/advisor roles, and some other compatibility areas are partial. Review the compatibility matrix and deployment guidance against your requirements before relying on those features.