SandBase Harness is a local-first runtime for AI agents that supplies the layer around the model loop: persistent sessions, sandboxed tool execution, memory stores, credential vaults, permission policies, audit trails, and a built-in Console for inspecting what an agent did. It stores metadata in SQLite and files on disk by default, with no hosted control plane, and runs on a developer machine or in self-managed infrastructure.
Its /v1 API follows the resource shapes of Claude Managed Agents, so the official Anthropic TypeScript SDK can drive it unchanged by pointing the client at the runtime. Compatibility is declared entry by entry in a capability matrix: of the official SDK's route surface, 76 routes are mounted, 29 refuse by name, and 5 (the multi-agent thread surface) are deferred. Agents can run on OpenAI, Anthropic, MiniMax, or any OpenAI-compatible provider, including DeepSeek V4, with one active provider configured at a time.
Generated code runs in a local process, per-session Docker containers, Kubernetes pods, or a self-hosted worker queue. Tool calls that need approval park the turn until a person or a preset policy answers, and sessions stream over resumable Server-Sent Events for replay and debugging. The runtime also ships a CLI, a TypeScript SDK, and a DeepSeek Harness plugin that exposes it over stdio MCP, plus Agent Plugins support for Copilot CLI and VS Code.
Features
- Managed Agents API: a /v1 API in Claude Managed Agents resource shapes that the official Anthropic SDK can call
- Sandbox backends: local process, per-session Docker containers, Kubernetes, and a self-hosted worker queue
- Persistent sessions: append-only event logs with resumable Server-Sent Events for replay and debugging
- Tool governance: MCP toolsets, built-in tools, custom tools, permission policies, and approval gates
- Credential vaults: credential injection, rotation, and OAuth refresh for tools and MCP servers
- Memory stores: versioned, audited memory that mounts into sessions and workers
- Session budgets: per-session cost ceilings priced from the event log
- Console: a local web UI for sessions, settings, and the API reference
- Model providers: OpenAI, Anthropic, MiniMax, and OpenAI-compatible endpoints such as DeepSeek
- Local storage: SQLite metadata and on-disk files for agents, sessions, environments, skills, and API keys
- Integrations: a DeepSeek Harness plugin over stdio MCP and Agent Plugins for Copilot CLI and VS Code
- SDK and CLI: a TypeScript SDK and a CLI for initializing workspaces, chatting with agents, and managing templates
