AGPL-3.0 自进化上下文数据库，统一智能体记忆、知识 RAG 与技能；通过 MCP 客户端 / Agent Plugins 通用集成接入 DeepSeek Harness，无原生 dsh 模式。
DSH 适配
生态相关
作者声明
安全审计
未审计
最后核验
2026-08-21
许可证
AGPL-3.0
01它能帮你完成什么？
Give an AI agent persistent memory, knowledge and skills in one browsable context database
Memories, resources and skills unified under the `viking://` protocol, browsable with ls/tree/find, L0/L1/L2 tiered on-demand loading, and observable retrieval trajectories
Developers building stateful AI agents who want deterministic, file-like access to agent memory instead of a black-box vector store
Wire long-term memory into a coding agent such as Claude Code, Codex, Cursor, MCP clients or LangChain
OpenViking recall injected into the agent's context and session memory auto-committed after each session
Coding-agent users who want cross-session memory and recall without manual note-taking
Run a ready-made AI agent on top of the context database
VikingBot agent started alongside the server via `openviking-server --with-bot`, chatted with via `ov chat`
Users who want an out-of-the-box agent built on OpenViking instead of integrating their own
02如何接入 DeepSeek Harness？
前置条件
- Python 3.10 or higher
- A model provider to configure during init — Volcengine, OpenAI, Codex OAuth, Kimi, GLM, or local Ollama
安装步骤
- 01
`pip install openviking --upgrade`
- 02
`openviking-server init` — interactive wizard: providers, models, writes `~/.openviking/ov.conf`
- 03
`openviking-server doctor` — validate setup
- 04
`openviking-server` — start the server (background: `nohup openviking-server > openviking.log 2>&1 &`)
验证接入成功
- Run `openviking-server doctor` — checks the config file, Python version, provider connectivity, and disk space without a running server
- With the server running, the bundled `ov` client CLI works, e.g. `ov status`
03DSH 适配与能力边界
No DeepSeek Harness-specific integration is documented in the README; OpenViking runs as a standalone context/memory server that agents reach through its generic integrations (MCP clients, Agent Plugins 1.0)
viking:// virtual filesystem for all context
memories, resources (docs, repos, web pages) and skills→one `viking://` URI tree browsed deterministically via `ov ls` / `tree` / `find` / `grep`
L0/L1/L2 tiered on-demand loading
any entry written into OpenViking→L0 abstract (~100 tokens), L1 overview (~2k tokens), L2 details — loaded only as deep as the task requires, cutting token spend
every entry is processed into three tiers on write (asynchronous semantic processing)Sessions become long-term memory
committed agent sessions→user preferences and agent experience extracted into long-term memory
after a session commits, OpenViking asynchronously extracts memory in the backgroundAgent integrations & ov CLI
Claude Code, Codex, OpenClaw, Hermes, Cursor, TRAE, OpenCode, MCP clients, LangChain/LangGraph and more→OpenViking recall injected into the agent's context with auto-committed session memory
`openviking-server init` writes `~/.openviking/ov.conf`; integrations modify the target agent's configuration
04适合谁？何时不该用？
适合
- Developers building stateful AI agents who want deterministic, file-like access to agent memory instead of a black-box vector store
- Coding-agent users who want cross-session memory and recall without manual note-taking
- Users who want an out-of-the-box agent built on OpenViking instead of integrating their own
不适合
- The server requires Python 3.10 or higher, plus a configured model provider (Volcengine, OpenAI, Codex OAuth, Kimi, GLM, or local Ollama) before it can serve context.
- The README documents no DeepSeek Harness (dsh)-specific integration; DSH users must rely on the generic integration lanes it lists (MCP clients, Agent Plugins 1.0), so a working dsh hookup is not guaranteed by the README alone.
- The project self-describes as early-stage, and the OpenViking Helper desktop console is still in beta for macOS and Windows x64 only.
05兼容性、维护与安全提示
- The server requires Python 3.10 or higher, plus a configured model provider (Volcengine, OpenAI, Codex OAuth, Kimi, GLM, or local Ollama) before it can serve context.
- The README documents no DeepSeek Harness (dsh)-specific integration; DSH users must rely on the generic integration lanes it lists (MCP clients, Agent Plugins 1.0), so a working dsh hookup is not guaranteed by the README alone.
- The project self-describes as early-stage, and the OpenViking Helper desktop console is still in beta for macOS and Windows x64 only.
AGPL-3.0 · actively maintained (latest release v0.4.15, 2026-08-18)
06常见问题
OpenViking 如何接入 DeepSeek Harness？
README 未提及 DeepSeek Harness 专属集成。OpenViking 提供通用接入通道——MCP 客户端和 Agent Plugins 1.0，另有 Claude Code、Codex、Cursor、TRAE、OpenCode 等专门指南，dsh 可走 MCP 客户端这一通用通道连接服务端。
是原生集成还是 MCP？
README 既未声明原生 dsh 支持，也未提供 dsh 插件；在其列出的集成方式中，MCP 客户端是通用的机器对机器接入通道，具体配置见其文档站，README 本身未展开。
安装前需要什么条件？
Python 3.10 及以上，并需在 `openviking-server init` 向导中配置模型提供方——火山引擎、OpenAI、Codex OAuth、Kimi、GLM 或本地 Ollama（向导可自动检测并安装 Ollama 运行时）。
我的数据会存在哪里？
自托管模式下数据保存在本地服务器，配置写入 `~/.openviking/ov.conf`；模型调用发往你配置的提供方（选本地 Ollama 则完全离线）。官方也提供托管在火山引擎上的可选 SaaS 版本。
有什么限制？
项目自述仍处于早期阶段；OpenViking Helper 桌面控制台仅以 Beta 形式提供 macOS 和 Windows x64 版本；生产使用建议按独立 HTTP 服务方式部署。
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。
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。
browserskill
作者 tencent
让 AI 智能体使用你已登录的真实浏览器而不打扰工作，CLI 与扩展双形态，支持任意可调用 shell 的智能体。
08数据与来源
Integrations inject OpenViking recall into your agent's context and auto-commit session memory:
- [MCP clients](https://docs.openviking.ai/en/agent-integrations/06-mcp-clients)
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。