Apache-2.0 纯 Markdown 自主科研系统：900+ 技能、四层层级与 7 个 MCP 集成；README 仅支持 Claude Code 与 Codex，无 DeepSeek Harness 路径（生态级）。
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
Apache-2.0
01它能帮你完成什么？
Turn a research direction into an executable, machine-readable Research Spec
A Research Spec with checkbox progress tracking, quantified completion criteria, backtrack conditions and session recovery, generated after North Star crystallization
Researchers and AI engineers who want an autonomous AI research agent rather than a search assistant
Run systematic literature acquisition and research-gap discovery
Multi-pass academic paper discovery via Semantic Scholar with citation chaining and snowball sampling, plus 15+ gap detection methods
Academic and industry researchers who need systematic literature coverage instead of keyword search
Continue long research campaigns across sessions without losing state
Checkpoint-based context files (context/INDEX.md + ≥500-line checkpoints) that let a new agent session resume from the spec's checkbox state
Teams running multi-session deep-research campaigns that span days or weeks
02如何接入 DeepSeek Harness？
前置条件
- Node.js 22 or newer
- Claude Code or Codex as the executing agent runtime — the two runtimes documented in the README
- API keys for the MCP servers you enable (Brave, Apify; Tavily optional; Semantic Scholar key optional — public API works keyless at lower rate limits)
安装步骤
- 01
Clone the repo and install dependencies: `git clone https://github.com/yogsoth-ai/de-anthropocentric-research-engine.git`, `cd de-anthropocentric-research-engine`, `npm install` (pulls the bundled MCP servers)
$ git clone https://github.com/yogsoth-ai/de-anthropocentric-research-engine.git
- 02
Codex: run `./install/codex.sh` (or `./install/codex.sh --target /path/to/your/project`; Windows: `.\install\codex.ps1`) — it adds a marked DARE section to the target's AGENTS.md and copies the knowledge base into `.dare/skills`; use `--dry-run` to preview
- 03
Claude Code: `cp mcp.example.json .mcp.json` and fill in your API keys
- 04
Claude Code: copy (or symlink) every skill into your project's `.claude/skills/` or the user-level `~/.claude/skills/` (e.g. `mkdir -p .claude/skills && cp -R skills/* .claude/skills/`)
- 05
Claude Code: invoke the entry point `/de-anthropocentric-research-engine`; later execute a generated spec with `/executing-specs`
- 06
Optional: for experiment-execution install the `superpowers` and `ponytail` Claude Code plugins via the plugin marketplace; for ara-from-context run `npx @ara-commons/ara-skills`
$ npx @ara-commons/ara-skills
验证接入成功
作者未说明
03DSH 适配与能力边界
Pure-markdown research-skill ecosystem; the README documents Claude Code (skills copied into `.claude/skills/`) and Codex (`install/codex.sh`) as the only runtimes and makes no DeepSeek Harness/dsh claim
Autonomous research orchestration (arsenal, not pipeline)
A research direction — cold-start from zero, warm-start from vague interest, or hot-start from a specific question→A crystallized North Star, then an executable Research Spec the agent executes stage-by-stage across 10 freely-composable packages, with explicit backtrack conditions
Writes Research Spec files (e.g. docs/de-anthropocentric/specs/) and context files during executionSystematic literature acquisition & gap discovery
A research topic or crystallized scope→Multi-pass paper discovery via Semantic Scholar (citation chaining, snowball sampling, cross-database verification) plus 15+ gap detection methods and 31+ ideation techniques
Network calls to Semantic Scholar and web-search MCP servers7 MCP integrations for external data access
Configured MCP servers: semantic-scholar, wiki-vault, brave-search, tavily-search, keenable, apify, alphaxiv→Paper lookup and citations, LLM-optimized web search, full-page scraping, and a persistent research knowledge graph (BM25 search, typed edges)
Requires API keys for some servers (BRAVE_API_KEY, APIFY_TOKEN; optional TAVILY_API_KEY, SS_API_KEY, VAULT_ROOT)keenable and alphaxiv connect to hosted remote endpoints (keyless public servers)Two Apify X-research Actors are paid and not loaded by the example configCheckpoint-based context management
A running research campaign spanning multiple sessions→Named context files with ≥500-line checkpoints, a context/INDEX.md registry, and automatic session recovery from the spec's checkbox state
Writes context files under context/ (gitignored at runtime)
04适合谁？何时不该用？
适合
- Researchers and AI engineers who want an autonomous AI research agent rather than a search assistant
- Academic and industry researchers who need systematic literature coverage instead of keyword search
- Teams running multi-session deep-research campaigns that span days or weeks
不适合
- DARE has no application code, runtime or framework — execution depends entirely on the host coding agent's ability to follow complex multi-step instructions; the README names Claude Code as the runtime, with Codex supported via the install script.
- The README documents no DeepSeek Harness (dsh) integration path — the Quick Start covers only Claude Code and Codex, so DSH users should treat DARE as an ecosystem research toolkit until a dsh path is published.
05兼容性、维护与安全提示
- DARE has no application code, runtime or framework — execution depends entirely on the host coding agent's ability to follow complex multi-step instructions; the README names Claude Code as the runtime, with Codex supported via the install script.
- The README documents no DeepSeek Harness (dsh) integration path — the Quick Start covers only Claude Code and Codex, so DSH users should treat DARE as an ecosystem research toolkit until a dsh path is published.
- Research campaigns reach external services via the 7 MCP servers — Brave and Apify require keys (Tavily and Semantic Scholar keys optional), while Keenable and AlphaXiv connect to hosted public endpoints; wiki-vault writes to a local vault directory you configure.
Apache-2.0 · actively maintained (latest release v3.2.2, 2026-06-21; last push 2026-08-15)
06常见问题
能在 DeepSeek Harness（dsh）里运行 DARE 吗？
按 README 现状不能。Quick Start 只覆盖两种运行时——Claude Code（把技能复制或软链到 `.claude/skills/`）和 Codex（`install/codex.sh` 在目标项目的 AGENTS.md 中写入 DARE 区块）。README 没有 dsh 安装路径，目前应把 DARE 当作生态级科研工具包。
DARE 是 MCP 服务器还是原生工具？
都不是——DARE 是纯技能架构：900+ 个 Markdown 指令文件由编码智能体直接读取并执行，没有应用代码和运行时。它*使用* 7 个 MCP 服务器（Semantic Scholar、Brave Search、Tavily、Keenable、AlphaXiv、Apify、Wiki Vault）获取外部数据。
需要什么前置条件？
Node.js 22 或更高版本、git，以及 Claude Code 或 Codex 运行时。`npm install` 会拉取内置的 MCP 服务器包。部分 MCP 服务器需要 API 密钥（Brave、Apify；Tavily 可选；Semantic Scholar 无密钥也能用，只是限流更低）；Keenable 和 AlphaXiv 是免密钥的托管公共端点。
我的研究数据会流向哪里？
MCP 调用会发往外部服务——Semantic Scholar、Brave/Tavily 网页搜索、Apify 抓取，以及托管的 Keenable 和 AlphaXiv 端点。Wiki Vault 把知识图谱写到你用 VAULT_ROOT 指定的本地 Obsidian 兼容库目录。两个 Apify X 数据抓取 Actor 是付费的，示例配置默认不加载。
主要限制有哪些？
执行完全依赖宿主智能体遵循复杂多步指令的能力。experiment-execution 包还需要额外安装 `superpowers` 和 `ponytail` 两个 Claude Code 插件，ara-from-context 需要外部的 `npx @ara-commons/ara-skills` 编译器。
07相关的 DSH 工作流
distilly
作者 titanwings
Distilly — Distill how they think into reusable Skills for any Agent or Bot. Formerly Colleague Skill（原同事 Skill）.
dsh-market
作者 dsh-market
The plugin market inside DeepSeek Harness — browse, search, one-click install · DSH 可视化插件市场
ai_animation
作者 unclecheng-li
本项目整理了用于生成[炫酷 HTML 动画网页]的 AI Prompts，涵盖动画效果、3D 可视化、PPT 风格演示、UI 美化等多个类别。
aegis
作者 ganyuanran
让 AI 编码智能体具备架构意识：基线优先、证据验证、漂移检查，保障长任务安全。
08数据与来源
The entire system is 900+ markdown files — each one a self-contained instruction set that Claude Code reads and executes…
Skills are auto-discovered from a `.claude/skills/` directory — there is no `settings.json` path option.
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。