sandbaseai/sandbase-harness
Open-source CMA-compatible agent runtime for any model, with MCP tools, sandboxed sessions, audit, replay, and
SandBase Harness is a local-first agent runtime that goes beyond the model loop — it handles persistent sessions, sandboxed tool execution, memory, credentials, audit trails, and a built-in Console. Think of it as the production infrastructure layer that agent SDKs don't provide: tool governance, sandbox boundaries (local, Docker, Kubernetes), credential vaults, resumable event streams for debugging and replay, and permission policies. A Claude Managed Agents-style /v1 API makes it compatible with existing agent clients. DeepSeek V4 is supported as a model provider; DSH sessions connect via MCP stdio bridge. The companion SandBase CLI bridges 25 AI client targets to 2,000+ models through a local stdio MCP server for simpler use cases. SQLite-backed: all agents, sessions, memory, and files live on your machine with no required hosted control plane.
Install
git clone --branch v0.3.7 --depth 1 https://github.com/sandbaseai/sandbase-harness.git && cd sandbase-harness && npm ci && npm run buildAfter building, create your agents workspace and start the server: mkdir ../my-agents && cd ../my-agents && node ../sandbase-harness/dist/index.js init && node ../sandbase-harness/dist/index.js start — then open http://127.0.0.1:3000/dashboard. Or try it instantly with GitHub Codespaces (see README). Also listed in the Official MCP Registry (io.github.sandbaseai/sandbase-harness).
Compatibility
Local-first: runs on your machine or own infrastructure. Supports: OpenAI, Anthropic, MiniMax, and any OpenAI-compatible provider including DeepSeek V4. Sandbox backends: local process, Docker (per-session containers), Kubernetes (kubectl exec/cp), self-hosted worker queue. DeepSeek Harness bridge available via MCP stdio.
Details
- Repo: sandbaseai/sandbase-harness
- Category: Agent Capabilities
- Stars: 629
- Version: v0.3.7 (see github.com/sandbaseai/sandbase-harness/releases/latest)
- Last push: 2026-09-21
- First seen: 2026-07-11
Recent updates
v0.3.7: Settings V2 (single workspace model vendor, sandbox, memory, storage config with validation and restart flow). Resumable SSE for session replay. MCP toolsets and permission policies. Listed on Official MCP Registry.
FAQ
- How do I start SandBase Harness?
- Clone the repo, build it (npm ci && npm run build), then: node dist/index.js init && node dist/index.js start — open http://127.0.0.1:3000/dashboard and configure a model in Settings > Models.
- Does SandBase Harness work with DeepSeek Harness?
- Yes — SandBase Harness has a DeepSeek Harness bridge over MCP stdio. Your DSH agents, sessions, and streamed turns connect through the MCP toolset layer.
- What is the difference between SandBase Harness and SandBase CLI?
- SandBase Harness is the full local agent runtime with sessions, sandboxes, memory, and Console. SandBase CLI is a lightweight MCP bridge connecting 25 AI client targets to 2,000+ models via local stdio — use CLI for simple bridging, Harness for production agent infrastructure.
Alternatives
deepseek-ai/deepseek-harness · Devin-AXIS/iPolloWork · mnemon-dev/mnemon