Apache-2.0 本地优先 Agent 运行时，以插件形式安装进 DeepSeek Harness：内置 MCP stdio 桥接，向 DSH 暴露沙箱会话、审计、回放与本地控制台。
DSH 适配
兼容
作者声明
安全审计
未审计
最后核验
2026-08-21
许可证
Apache-2.0
01它能帮你完成什么？
Run sandboxed, auditable agent sessions from inside DeepSeek Harness
DSH can list agents, create and run sessions, stream turns, inspect results and artifacts, and stop work through native mcp__sandbase__* tools
DSH users who want a self-hosted agent runtime with sandboxing, audit trails and replay behind their DSH profile
Execute generated agent code safely in isolated sandboxes
Session sandboxes on local process, Docker (per-session containers), Kubernetes (kubectl exec/cp), or a self-hosted worker queue, plus audit trails and replay
Teams running untrusted or generated code who need sandbox boundaries, credential handling and auditability
Operate any model provider, including DeepSeek V4, from one local-first runtime
A single configured model provider boundary (OpenAI, Anthropic, MiniMax, or OpenAI-compatible endpoints) driving agents and sessions, with SQLite metadata kept local
Developers who want model-agnostic agents without a required hosted control plane
02如何接入 DeepSeek Harness？
前置条件
- DeepSeek Harness (dsh) installed with a profile available (the README example uses `--profile web`)
- Node.js 22+ and npm 10+
- A model provider API key (OpenAI, Anthropic, MiniMax, or an OpenAI-compatible endpoint)
- Docker (optional, for Docker-backed sandboxes)
安装步骤
- 01
Clone the tagged release and build: `git clone --branch v0.3.7 --depth 1 https://github.com/sandbaseai/sandbase-harness.git`, `cd sandbase-harness`, `npm ci`, `npm run build`
$ git clone --branch v0.3.7 --depth 1 https://github.com/sandbaseai/sandbase-harness.git
- 02
Create a workspace and start the runtime: `mkdir ../my-agents && cd ../my-agents`, `node ../sandbase-harness/dist/index.js init`, `node ../sandbase-harness/dist/index.js start`
- 03
Point DSH at the running runtime: `export MANAGED_AGENTS_URL=http://127.0.0.1:3000`
- 04
Install the bundle into a DSH profile and boot it: `dsh plugin --profile web add -w ../sandbase-harness`, then `dsh web`
$ dsh plugin --profile web add -w ../sandbase-harness
验证接入成功
- Open `http://127.0.0.1:3000/dashboard`, go to Settings > Models, paste your API key — the README says "you're running"
03DSH 适配与能力边界
Installed into a DSH profile as a third-party plugin (`dsh plugin --profile web add -w ../sandbase-harness`); a bundled MCP bridge over stdio exposes agents, sessions, streamed turns, artifacts and cancellation as native mcp__sandbase__* tools
DeepSeek Harness MCP bridge
a running SandBase Harness runtime reachable at MANAGED_AGENTS_URL→agents, sessions, streamed turns, artifacts and cancellation exposed to DSH as native mcp__sandbase__* tools over stdio
Sandboxed session execution
agent sessions with generated code or tool calls→isolated execution on local process, Docker (per-session containers), Kubernetes (kubectl exec/cp), or self-hosted worker queue
Docker-backed mode starts per-session containers; local mode writes session sandboxes and workspace snapshots to the .managed-agents state directoryPersistent sessions with audit and replay
running agent sessions→persistent sessions, resumable Server-Sent Events for replay and debugging, and audit trails stored in SQLite
writes SQLite metadata (data.db), logs, uploaded files and skill bytes to the workspace state directoryMulti-provider model operation
one configured provider boundary via Settings (OpenAI, Anthropic, MiniMax, or OpenAI-compatible endpoints, including DeepSeek V4)→agents and sessions served against the chosen model provider
model requests go over the network to the configured provider; an API key is required
04适合谁？何时不该用？
适合
- DSH users who want a self-hosted agent runtime with sandboxing, audit trails and replay behind their DSH profile
- Teams running untrusted or generated code who need sandbox boundaries, credential handling and auditability
- Developers who want model-agnostic agents without a required hosted control plane
不适合
- The unscoped `managed-agents` name on npm is not this project. Install only from the tagged GitHub source release; do not run `npx managed-agents` or `npm install managed-agents`.
- The runtime API is open by default; authentication only activates once at least one API key exists (static MANAGED_AGENTS_API_KEY or a managed key), so keys should be configured before exposing the port beyond localhost.
05兼容性、维护与安全提示
- The unscoped `managed-agents` name on npm is not this project. Install only from the tagged GitHub source release; do not run `npx managed-agents` or `npm install managed-agents`.
- The runtime API is open by default; authentication only activates once at least one API key exists (static MANAGED_AGENTS_API_KEY or a managed key), so keys should be configured before exposing the port beyond localhost.
- Requires Node.js 22+, npm 10+ and a model provider API key to run; Docker is additionally required for Docker-backed sandboxes, and Codespaces usage may be billed by GitHub.
Apache-2.0 · actively maintained (latest release v0.3.7, 2026-08-20)
06常见问题
如何把 SandBase Harness 安装到 DeepSeek Harness？
克隆带 tag 的发布版本，用 npm 构建，在工作区 init 并启动运行时，设置 MANAGED_AGENTS_URL=http://127.0.0.1:3000，然后执行 `dsh plugin --profile web add -w ../sandbase-harness` 和 `dsh web`。DSH profile 直接安装经过校验的源码检出，而不是 npm 包。
这个集成是原生的还是走 MCP？
它是 DSH 插件：安装 bundle 后 DSH 会通过 stdio 启动内置的 MCP 入口，把 Agent、会话、流式回合、产物和取消操作暴露为原生 mcp__sandbase__* 工具。
安装前需要准备什么？
Node.js 22+、npm 10+、一个模型服务商 API 密钥（OpenAI、Anthropic、MiniMax 或 OpenAI 兼容端点），如需 Docker 沙箱还要安装 Docker。运行时必须已启动，保证 MANAGED_AGENTS_URL 可访问。
我的数据会存在哪里？
本地优先：SQLite 元数据、上传文件、技能包、快照和会话沙箱都保存在工作区的 `.managed-agents` 状态目录中；模型请求发往你自行配置的模型服务商，不依赖任何托管控制平面。
需要 SandBase 账号或 npm 包吗？
当 DSH 已提供网页/搜索工具时无需 SandBase 账号。另外 npm 上未加 scope 的 `managed-agents` 名称并非本项目——只应从带 tag 的 GitHub 源码发布安装。
07相关的 DSH 工作流
archify
作者 tt-a1i
为编码智能体生成美观可验证的架构图、时序图与数据流图，输出自包含 HTML，支持动效与清晰导出。
openviking
作者 volcengine
为 AI 智能体打造的自进化上下文数据库，统一智能体记忆、知识 RAG 与技能。
nocobase
作者 nocobase
开源的 AI + 无代码应用搭建平台：AI 在久经生产验证的底层基础设施与所见即所得的无代码界面之上工作，帮你快速构建 CRM、ERP 等业务系统，兼顾速度与可靠性。
learn-harness-engineering
作者 walkinglabs
Harness 工程新手教程，从 0 到 1 系统学习智能体工作流框架。
08数据与来源
Run this project as a DSH plugin instead of treating `dsh-plugin` as discovery metadata only. Install the bundle into a…
DeepSeek Harness bridge over MCP stdio for agents, sessions, streamed turns, artifacts, and cancellation
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-08-21。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。