Polaris：用 DeepSeek Harness 生态做自主 AI 科研

Apache-2.0 自主 AI 科研平台（从文献到完成评审的论文），通过 Docker Compose 部署；其只读 MCP 服务器可作为 DeepSeek Harness 生态伙伴接入编码智能体。

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？

Run the full research lifecycle — literature survey, idea generation and review, GPU experiments, LaTeX paper writing, and paper review — as one platform
Cross-linked research wiki, scored and deduplicated ideas with Elo-ranked review, experiment runs with streamed logs and metric curves, multi-file LaTeX papers with live PDF preview, and citation-verified reviews
Research labs and AI research teams that want an autonomous AI-scientist platform covering the whole pipeline
f97a1b92aa78…f97a1b92aa78…
Let agents run long experiments on lab GPU servers with self-verification and human gates
Automated experiment Voyages: study plan, compute-budget check, generated code, smoke test, streamed runs with live metric curves, and auto-iteration that repairs failures under a time budget
ML researchers who want agents to execute and iterate on real GPU servers under budgets and audit
f97a1b92aa78…
Expose Polaris's research knowledge to external coding agents via MCP
An MCP server (Streamable HTTP and stdio) of read-only tools over literature, knowledge, project state, manuscripts, and external search
Users of coding agents (Claude Code, Codex, Cursor) who want research context inside their agent workflow
f97a1b92aa78…

02如何接入 DeepSeek Harness？

前置条件

Docker and Docker Compose only — the README says no local Python, Node, or database is needed for the recommended Docker Compose setup
LLM provider keys and secrets configured in .env

安装步骤

01
Run `cp .env.example .env` and set provider keys and secrets
02
Run `make dev` to start the full stack via docker compose with hot reload (frontend at http://localhost:5173, backend API docs at http://localhost:8000/docs)
03
Production alternative: deploy pre-built images — set POLARIS_ENV=prod and POLARIS_IMAGE_TAG in .env, run `docker compose --env-file .env -f docker/docker-compose.yml pull`, then `up -d`, then `docker compose -f docker/docker-compose.yml exec api alembic upgrade head` (required on first run)

验证接入成功

作者未说明

03DSH 适配与能力边界

DSH 适配生态相关

DSH-ecosystem AI research platform (tagged dsh-plugin on GitHub): run Polaris as a standalone web app via Docker Compose; its MCP server exposes read-only research tools to external coding agents

Voyage agent core
any long-running research task (literature backfill, experiment run, paper drafting)→resumable, auditable runs with a persisted planning → executing → verifying loop that resumes from checkpoint after a crash
retains every plan, action, and verdict, replayable in the UI; auto-pauses when budgets are exceeded
f97a1b92aa78…
Research Wiki (compile, don't retrieve)
papers from OpenAlex, Semantic Scholar, and arXiv→one cross-linked wiki per paper shared platform-wide (TL;DR, method, reusable ideas, concept backlinks), research digests, and Obsidian vault sync
queries external literature APIs (OpenAlex, Semantic Scholar, arXiv) and writes wiki/Obsidian artifacts
f97a1b92aa78…
Experiment Lab over SSH
per-user, Fernet-encrypted SSH credentials to the lab's GPU servers→planned, budget-checked experiment runs with streamed logs, live metric curves, generated figures, and a file-based memory kept across steps
writes and runs code on remote GPU servers under gated remote writes, command allow/deny lists, and triple budget caps
f97a1b92aa78…
MCP tool layer
Polaris project data (literature, knowledge, project state, manuscripts)→an MCP server (Streamable HTTP and stdio) exposing read-only tools to Claude Code, Codex, and Cursor, with a self-check playground
exposes project data over the network (Streamable HTTP / stdio); strictly read-only and project-isolated
f97a1b92aa78…

04适合谁？何时不该用？

适合

Research labs and AI research teams that want an autonomous AI-scientist platform covering the whole pipeline
ML researchers who want agents to execute and iterate on real GPU servers under budgets and audit
Users of coding agents (Claude Code, Codex, Cursor) who want research context inside their agent workflow

不适合

The README does not document running inside DeepSeek Harness (dsh): Polaris is deployed as a standalone web app via Docker Compose; its connection point to external agents is a read-only MCP server documented for Claude Code, Codex, and Cursor.
The recommended deployment is a full multi-container stack (PostgreSQL 16 with pgvector, Redis 7, API, worker, frontend), and the first-run `alembic upgrade head` migration is mandatory since Postgres tables are not auto-created.

05兼容性、维护与安全提示

The README does not document running inside DeepSeek Harness (dsh): Polaris is deployed as a standalone web app via Docker Compose; its connection point to external agents is a read-only MCP server documented for Claude Code, Codex, and Cursor.
The recommended deployment is a full multi-container stack (PostgreSQL 16 with pgvector, Redis 7, API, worker, frontend), and the first-run `alembic upgrade head` migration is mandatory since Postgres tables are not auto-created.
Desktop builds are neither signed nor notarized: macOS needs `xattr -dr com.apple.quarantine`, Windows SmartScreen requires Run anyway, and the Linux AppImage needs extra system libraries.

2026-07-222026-08-18v0.3.8

Apache-2.0 · actively maintained (latest release v0.3.8, 2026-08-17)

06常见问题

Polaris 如何与 DeepSeek Harness 对接？

README 未提供 dsh 原生集成说明。Polaris 以独立的 Web 应用运行，通过 Docker Compose 部署；它对外部编码智能体的接入点是只读 MCP 服务器（Streamable HTTP 和 stdio），文档中以 Claude Code、Codex 和 Cursor 为例。

运行 Polaris 需要准备什么？

推荐方式只需 Docker 和 Docker Compose，无需本地安装 Python、Node 或数据库。将 `.env.example` 复制为 `.env`，填入 LLM 提供商密钥和机密信息，然后运行 `make dev`（生产环境可部署 Docker Hub 上的预构建镜像）。

智能体能在真实 GPU 服务器上跑实验吗？

可以。实验中心使用按用户 Fernet 加密的 SSH 凭据，先规划研究方案、通过算力预算检查，再编写并运行代码，提供流式日志与实时指标曲线，并在受控的远程写入、命令允许/拒绝列表和三重预算上限下自动迭代。

Voyage 是什么？

每个长任务都是一次 Voyage：由 Navigator（规划）、Helm（执行）、Sextant（自我验证）三段循环驱动的持久化、可恢复、人工把关的智能体运行。工作进程崩溃后可从检查点续跑，预算超限时会自动暂停。

桌面客户端安装要注意什么？

构建未经签名和公证，各系统首次运行需放行一次：macOS 执行 `xattr -dr com.apple.quarantine /Applications/Polaris.app`，Windows 在 SmartScreen 中选择「更多信息 → 仍要运行」，Linux AppImage 需要 libnss3、libgtk-3-0 和 libasound2。

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。

精选Agent、自动化与工作流Python

30.8K2,378

colleague-skill

作者 titanwings

将冰冷的离别化为温暖的 Skill，欢迎加入数字生命1.0！Transforming cold farewells into warm skills? It's giving rebirth era. Welcome to Digital Life 1.0. 🫶

Agent、自动化与工作流Python

23.6K2,095

archify

作者 tt-a1i