中文版

LucentThinker（GitHub：https://github.com/XinYu-pumch/LucentThinker ，2026-10-31 创建，GPL-3.0）是一个本地运行的开源深度研究助手：给它一个问题，它会自己上网搜索、阅读原文，最后写出一份带引用的研究报告。可以理解为 GPT Deep Research 的开源本地版，动力来自 DeepSeek，自备 API Key。项目刚发布几天、star 还不多，但完成度和想法都在线。

它能干什么

• 深度研究全流程：提问之后自动多轮搜索、抓取并阅读来源，产出带出处引用的研究报告，而不是一段没有来源的聊天回答
• 深度可调、记录在本地：研究深度可以选，历史记录保存在本机，不用担心调研内容外流
• 两种用法：独立 Web UI（本机 localhost:8765），或者装成 DeepSeek Harness（dsh）的插件（插件名 dsh-plugin-lucentthinker，带 macOS Apple Silicon 与 Windows x64 运行时）
• 成本低：模型与搜索抓取都填 DeepSeek 官方 Key 即可，作者给出的参考成本是单次研究约 1 元人民币起（视任务大小和服务商价格浮动）

怎么用

• 独立版：装好 uv，clone 仓库，macOS 跑 ./start.sh --open（或双击 Launch LucentThinker.command），Windows 双击 start.cmd，依赖自动安装，界面在 localhost:8765 打开
• 配置：在设置里把 DeepSeek API Key 填到 DeepSeek 与 Search & Fetch 两处，刷新模型列表后添加 DeepSeek Flash，思考等级设为 Max，并设为主模型与抽取模型；也支持其他模型服务商
• 插件版：在 dsh 里打开 Plugins → Add plugin，输入 dsh-plugin-lucentthinker 启用即可

关于它的 benchmark

• 作者用能源、健康、软件、商业、贸易 5 个任务，对比了 GPT Deep Research 与付费版 MiroThinker，称研究质量评分更高；但这是作者自行组织的非盲测，每个产品每个任务只有一份回答，样本很小，结果参考着看即可

适合谁

经常要做调研的人：技术选型、行业摸底、面试前查公司，花大约 1 元就能跑一份带引用的报告，比订阅制的深度研究工具划算得多，记录还都留在本地。

English Version

LucentThinker (GitHub: https://github.com/XinYu-pumch/LucentThinker , created 2026-10-31, GPL-3.0) is a locally run, open-source deep research assistant: give it a question and it searches the web, reads the sources itself, and writes a research report with citations. Think of it as an open-source, local GPT Deep Research, powered by DeepSeek with your own API key. The project is only a few days old with few stars so far, but the idea and the polish are both there.

What it does

• The full deep-research flow: after you ask, it runs multiple rounds of search, fetches and reads sources, and produces a report with citations — not a sourceless chat answer
• Adjustable depth, local history: research depth is configurable, and history is stored on your own machine
• Two ways to use it: a standalone Web UI (localhost:8765 on your machine), or as a plugin for DeepSeek Harness (dsh) — plugin name dsh-plugin-lucentthinker, with runtimes for macOS Apple Silicon and Windows x64
• Cheap to run: one official DeepSeek API key covers both the model and search/fetch; the author’s reference cost starts around RMB 1 (roughly US$0.20) per run, depending on task size and provider pricing

How to run it

• Standalone: install uv, clone the repo, then on macOS run ./start.sh --open (or double-click Launch LucentThinker.command); on Windows double-click start.cmd. Dependencies install automatically and the UI opens at localhost:8765
• Configure: in Settings, enter your DeepSeek API key under both DeepSeek and Search & Fetch, refresh the model list and add DeepSeek Flash, set its thinking level to Max, and select it as both the main model and the extraction model; other providers are supported too
• Plugin: in dsh, open Plugins → Add plugin, enter dsh-plugin-lucentthinker, and enable it

About its benchmark

• The author compared it against GPT Deep Research and the paid MiroThinker on five tasks spanning energy, health, software, business and trade, and reports higher research-quality scores; note, however, that this is an author-organized, non-blind evaluation with a single response per product per task — a very small sample, so treat the result as a reference only

Who it’s for

Anyone who researches things often: technology choices, industry briefings, or background checks before an interview. For roughly a yuan you get a cited report, far cheaper than subscription deep-research tools, with all records kept locally.