2026-11-12 · GitHub AI 日报

统计窗口2026-11-11 10:30 ~ 2026-11-12 10:30（北京时间），对应 GitHub `committer-date >= 2026-11-11T02:30:00Z` 之后的默认分支提交。本期共收录 **273** 个条目，其中精选 5 条，四类分布为 ① 119 / ② 49 / ③ 45 / ④ 60。

覆盖方向AI Coding / 编程智能体与技能生态 · 医学 / 临床研究 AI 与 Agent 技能库 · 健康管理 / 营养与食品 AI · 模型底座与 Agent 基础设施

去重说明已与 10-16 至 11-11 二十三期推送记录逐条比对（用 `grep -o "https://github.com/[^/]*/[^/]*"` 抽出**累计 1548 个历史仓库全集**再逐一比对，且比对时统一转小写以免大小写变体漏网），本期 273 个条目**全部为首次出现，零重复**；5 条精选亦均为首次入榜。本轮检索共采集 **668** 个 `search/repositories` 唯一候选（30 条单词查询配合 `pushed:>2026-11-10` 日期限定），另单独抓取 GitHub trending 日榜 / 周榜 / 月榜共 **39** 个仓库逐个核实；与历史重复的 113 个先剔除，余 555 个非历史候选再**逐仓用 core 端点 `GET /repos/{o}/{r}/commits?since=&per_page=1` 核实默认分支提交**（读取 `Link` 头 `rel="last"` 页码作为精确提交数），其中 **523 个在窗口内有真实提交**、**36 个窗口内零提交直接剔除**，人工挑选与信息密度筛选后定稿 273 个。**凡 `pushed_at` 显示更新但逐仓核实后落在窗口之外的仓库一律剔除，不凭 `pushed_at` 收录；trending 榜在列也一律单独核实后才决定收录。**

核验口径所有"今天有动作"条目均以 GitHub 提交记录核实，**精确提交数一律以 core 端点 `Link` 头 `rel="last"` 的页码为准**（`per_page=1` 时页码即总条数）。**星标与许可证取自 GitHub API 实时快照**（截至 2026-11-12 10:30 北京时间）。**许可证以 core 端点 `GET /repos/{o}/{r}/license` 读取的 LICENSE 文件实文为准，API 的 `spdx_id` 只作线索**——本期实测 `use-junction/usejunction` 的 API 报 `NOASSERTION`，实文是**专有授权**；`awslabs/hcls-agent-skills` 的 API 报 `MIT-0`，实文是 **Amazon 版权声明 + MIT-0 式授权**（两者一致）；另有 **86** 个条目经 core 端点逐一确认**仓库根目录没有 LICENSE 文件**，表中标注为"未声明"——**"未声明"不等于无授权条款，但选型阶段必须回查**。本期 273 个条目中 MIT 与 Apache-2.0 合计 170 个，加 BSD / ISC / MPL / CC0 等宽松型授权后共 174 个；AGPL-3.0 与 GPL / LGPL 合计 11 个属于传染性 copyleft，只能读不能抄；ELv2 / BUSL-1.1 / PolyForm NC 等"源码可见但非开源" 1 个；专有授权 1 个；另有 86 个无 LICENSE 文件。选型时按三档分流。

一、今日精选 5 条

1. awslabs/hcls-agent-skills：AWS 把 43 个医疗与生命科学技能做成了「可移植的标准件」，并用 410 道两两对比题证明它比不加载技能更强

地址：https://github.com/awslabs/hcls-agent-skills ｜ ★76 ｜ MIT-0 ｜ Open-source healthcare AI agent skills and life sciences agent skills. 42 skills for medical coding, risk adjustment, claims analytics, quality measures, genomics, imaging, and drug discovery across 13 HCLS domains.

事件内容

本期最贴我业务的一条：这是目前公开可查的、由大厂维护的「医学 / 临床 Agent 技能库」，而且它自带一份像样的评测报告。 README 给出 43 个技能、覆盖 13~14 个领域，并按两类划分：reasoning 技能编码方法论与判定框架（如 ACMG 变异分级、目标试验仿真、HEDIS 指标定义），pipeline 技能编码可直接运行的工具链（GATK4 变异检测、RDKit 化学信息学、HL7v2 解析）。
它用一份 410 题的两两对比评测（380 道单技能 + 30 道跨技能）证明技能本身有用：由 Claude Opus 4.7 当评审，技能版在头对头对比中胜出 70~86%（视 harness 而定），批判性思维维度最高 85%、科学准确性维度最高 87%。技能遵循 Agent Skills 开放标准，可在 20+ 个平台上跑（Kiro、Amazon Quick、Bedrock AgentCore、Strands SDK、Claude Code、OpenAI Codex、CrewAI、LangChain 等）。
窗口内的两条提交正好落在「评测可信度」上：fix(eval): Extend fail-loudly to independent judge, fix stale docs 把「平局当失败」的原则从本地评审扩到独立评审模型；另一条合入 roda-open-data 技能（ODAC 开放数据访问规范）。
许可证是 MIT-0，仓库根目录 LICENSE 实文为 Amazon 版权声明 + MIT-0 式授权（API 返回的 spdx_id 亦为 MIT-0），选型上属于最宽松的一档。

值得关注的原因

这是「把专家经验写成可移植资产」这件事的工业级样本，正好是我专家包工作流缺的那个模板。 我现在的专家包质检还是「让模型自评」，而 hcls-agent-skills 的做法是：技能写成带明确判定阈值的文件（ACMG 分级、HEDIS 指标），再用两两对比 + 独立评审模型去测。它的 reasoning 与 pipeline 二分法我可以直接照搬——我的 4P 精准营养报告里，「解读准则」（线粒体、甲基化、OAT 的判定顺序）属于 reasoning，「DICOM/化验指标解析与导出」属于 pipeline。
「技能可跨 20+ 平台且遵循开放标准」这一点对我的一人公司极其关键。 我的东西不能绑死在某一个客户端上；标准化意味着同一套规则文件可以在 Claude Code、Codex、本地 dsh 之间复用，不必为每个 harness 重写一遍。
它印证了我 11-08 到 11-11 连续四期记下的那条主线，且给出了评测侧的完整做法。 我记过 assay 的「读声明必须配文件清单」、groundedmed 的「批评者模型逐条对照原文」、bpf-benchmark 的「第三方独立评测制品」——这一期 hcls-agent-skills 把三者合到了一份可下载的资产清单里：技能本身是方法论，独立评审是裁判，两两对比是量表。如果我给 4P 专家包做一套 30~50 题的两两对比评测，方法论可以直接照抄这一套。
技能应当是「知识分子的决策链」而不是「长提示词」，这个论点对我做内容有直接价值。 README 那句「模型知道事实但不知道程序」正好解释了我公众号科普反复遇到的困境：读者需要的是「下一步该做什么判断」，不是概念列表。

2. openai/codex-security：官方给 agent 装上了「安全扫描器」，并且开始把别的扫描器的结果也纳入统一证据链

地址：https://github.com/openai/codex-security ｜ ★11,802 ｜ Apache-2.0 ｜ OpenAI's Codex Security CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities. npm: https://www.npmjs.com/package/@openai/codex-security

事件内容

OpenAI 官方的 Codex Security CLI 与 TypeScript SDK，用于在代码里「发现、验证并修复漏洞」；窗口内共 34 次提交，并在 2026-11-12 发布 0.10.0（npm 包与 GitHub Action 同时发版）。
本期最有意思的一条是「把第三方扫描器的结果收编进同一套证据链」：先是 feat(scanner-imports): add ArmorCode findings support，随后 feat(scanner-imports): import AWS Continuum code issues，再是 feat(import): include AWS-generated patches in Continuum findings——AWS 那边生成的补丁，会被原样带进 Continuum 的发现条目里。
另一条是「可恢复」而非「一次性」：feat(scans): resume interrupted Standard scans 让被中断的 Standard 扫描能接着跑；fix(diagnostics): distinguish configuration errors and policy refusals 把「配置错了」和「策略拒绝」分成两类不同的诊断，不再混成一锅。
它支持导出 SARIF / JSON / CSV、推送到 Linear / Jira / GitHub，也能起草 SECURITY.md 策略并把威胁模型存档待审。

值得关注的原因

「AWS 生成的补丁被带进第三方扫描器的发现条目」这件事，方法论上比功能本身更值钱。 这是一条完整的证据链：发现（AWS Continuum）→ 补丁来源（AWS 生成）→ 归属（第三方扫描器条目）→ 导出格式（SARIF/JSON/CSV）。这正是我 GRPM 报告最缺的那一段——11-10 我记 bpf-benchmark 的「哈希绑定证据」、11-11 我记 groundedmed 的「判定规则先于运行落盘」，都属于「结论怎么来的」，而这条链讲的是「修改怎么来的」。对我给客户出的干预报告来说，「建议动作」同样需要绑定来源与生成方式，否则三个月后无法复核。
「区分配置错误与策略拒绝」是一条能立刻抄进我专家包质检的错误分类法。 我的报告里如果某个建议没给出，读者分不清是「规则没覆盖到」还是「规则判定为不适用」——这两种错误必须分开报，否则用户会以为系统坏了。
「可恢复的扫描」对应我一人公司每周 8 小时的现实约束。 一个跑两小时就断了的任务等于没有；能中断、能续跑，意味着我可以在通勤时启动、回来接着用。这条对所有长任务型 agent 都成立，值得写进我的技术选型清单。

3. awslabs/unified-kg-rag-on-aws：四种图谱 RAG 跑在同一套索引上，并且开始把「重命名」当成一次需要回滚的迁移

地址：https://github.com/awslabs/unified-kg-rag-on-aws ｜ ★39 ｜ Apache-2.0 ｜ AWS-native knowledge graph RAG framework. Runs GraphRAG, LightRAG, HippoRAG 2 and RAPTOR on one shared index, each selectable at index and query time, with hybrid BM25 + vector + graph retrieval and source provenance. Amazon Bedrock, Neptune and OpenSearch, or fully local. Incremental indexing, multilingual, bring-your-own KG, built-in evaluation.

事件内容

在同一套索引上同时跑四种已发表方法：Microsoft GraphRAG（社区摘要）、LightRAG（双层关键词）、HippoRAG 2（实体与段落上的 Personalized PageRank）与 RAPTOR（摘要树，需显式开启）——共用摄入、索引、缓存与检索，只有召回算法不同，而且只构建被选中方法真正要读的部分。
答案带 [n] 引用标记，指名背后的文档与 PDF 页码（global 与 drift 两种策略默认不带）；点开来源可以看到「溯源路径」一直走到力导向图里被圈出的实体及其邻居。检索层是 BM25 + 向量 + Neptune 图遍历，用 RRF 融合并由 Bedrock 重排模型重排。
窗口内的三条提交全在处理「实体重命名」：先是 fix: re-key bundle — name folding and legal forms, prompt-aware stage cache keys，再是 fix: legal forms read as written, idempotent name keys, SUSPECTED claims, prompt digests，最后是 fix(indexing): strip a vertex holding a kept edge, identifier tokens from the entity key, record ids mint。核心是「改名要幂等」——同一个实体的新旧名字必须映射到同一把键，且被重命名后仍在被别的边引用的顶点不能被顺手删掉。
评测默认不依赖 LLM 裁判：内置 answer match、retrieval hit rate、graph coverage 三项无 LLM 指标，可选 RAGAS 与断言式评审，并给出成对区间比较与每次查询的成本计量。

值得关注的原因

「重命名是一次需要回滚的迁移」这句话直接命中我 GRPM 的术语漂移问题。 我 11-10 记过 cortexdb 的解法是「别名在写入和读取时都能解析」，而这里给出了另一条更狠的路：重命名不是加一个别名，是要把旧键的边全部迁走、迁不走的顶点必须留住、而且必须幂等（同一实体改名两次不能产生两个身份）。 我的术语体系升级（把「omega-3」和「EPA+DHA」合并、给新指标换名）迟早要动历史结论，现在就需要知道别人是怎么处理「改名后旧结论还算不算数」的。
「一个被别的边引用的顶点不能被顺手删掉」是一条我在内容侧从没想过、但很关键的规则。 我做公众号 A/B 测试时会改标题、换术语，如果旧标题对应的结论被直接删掉，历史测试就再也复现不了——这跟数据侧的删顶点是同一个道理。
「默认不用 LLM 裁判」这条方向与我这几期的判断完全一致。 我连续四期记的都是「让一个不知情的第三方来判」和「机器校验优先于模型自述」，而这里更进一步：三项无 LLM 指标作为默认，LLM 裁判作为可选开关。 这个默认值的选择，比它的功能列表更值得抄。
成本计量 + 成对区间比较是内容实验该有的样子。 我的公众号内容测试目前只有「这篇阅读量高」这种粗口径；「每种策略配对区间 + 每次查询成本」这个结构，可以直接改造成「两版文案 A/B 的成对区间 + 生成成本」的模板。

4. CaiZongyuan/clinmesh：把中国公立医院的挂号—医嘱—药房—医保—病案做成可评测的虚拟环境，而且明确「最后落笔的必须是人」

地址：https://github.com/CaiZongyuan/clinmesh ｜ ★3 ｜未声明｜面向 Agent 评测、产品验证和医疗信息系统研究的中国公立医院 HIS 虚拟环境，提供 FHIR R5 接口与可重复、可审计的合成业务场景。

事件内容

面向 Agent 评测、产品验证与医疗信息系统研究的中国公立医院仿真 HIS，以 FHIR R5 为标准互操作接口，用显式业务命令承载挂号、医嘱、药房、收费、医保、库存与病案等流程。仓库自述只处理虚构、合成或不可逆脱敏的演示数据。
权限边界写得很清楚：ClinMesh Agent Tools 只能读取当前受信页面上下文、导航、编辑草稿与准备人工审阅；「正式挂号、分诊、临床、支付和药房 Command 的最终 Actor，始终是点击原生审阅框的登录人类」。Agent 无法直接提交业务动作。
FHIR 公开面固定为 R5 5.0.0，只声明并实现 resource read、vread、instance history 与白名单 Search；业务写入只走受控 Command，服务器不宣告通用 FHIR create/update/delete，也不宣告自定义 Profile 或 Implementation Guide 一致性。
窗口内三条提交：把 GitHub Issue 恢复为正式需求规格并退役 docs/spec 目录、把 spec 标签收敛回规格主票并排除执行子票、集成临床工作台连续办理。需求规格从文档目录回到 Issue 本身——这是流程上的一次「回到源头」。

值得关注的原因

这个项目最值得我看的不是 FHIR 实现，而是它对「agent 能碰到什么」的边界划法。 它没有靠提示词说「请不要替医生做决定」，而是在架构上让 Agent 拿不到提交按钮——写动作只能通过人点审阅框完成。这比我给自己定的「不宣称医疗资质」更硬：我的红线目前只是文案层面的自律，而 ClinMesh 是把它做成了系统层的不可绕过。 对我的 4P 平台同理——如果哪条建议「必须经人确认」，就该做成流程上的必经步骤，而不是提示词里的一句提醒。
「合成数据 + 可重复 + 可审计」这三条同时写进定位，对我百十工程和真实世界研究有直接参考价值。 我的数据脱敏流程一直是个软性问题（11-09 记过 dsh-medical-anonymizer 是可选项而非必需），而 ClinMesh 的做法是从一开始就不接入真实数据，而不是事后清洗。
「需求规格从文档目录回到 Issue」这条对我有立刻可用的启发。 我 4P 平台的实施文档体系一直存在「文档一套、实际一套」的漂移风险——用可勾选的 Issue 当规格唯一真相源，文档只做渲染，这条能直接抄。
它同时接了 DSH（DeepSeek Harness）的原生 React Surface。 我本机装着 dsh，这条线已连续三期（11-09 dsh-Go、11-10 dsh-medical-anonymizer、11-11 dsh-plugin-radar、12 期 dsh-context）出现在我的日报里，说明 dsh 生态正在从「个人插件」走向「行业场景适配」，这个趋势值得单独跟踪。

5. rupal2k/openfit：一天发三个版本，把自托管健身 App 的体检报告 OCR 做成了「读完即删」的隐私边界，并公开了自己 95 条规则的验证状态

地址：https://github.com/rupal2k/openfit ｜ ★1 ｜ AGPL-3.0 ｜ Self-hosted training, body-weight and nutrition tracker with on-device AI vision: guided workouts, progressive overload, passkey sync, PWA and standalone Android.

事件内容

自托管的训练、体重与营养追踪应用，一个 Docker Compose 栈跑在用户自己的服务器或 homelab 上，不连外部数据库；支持引导式训练与渐进超负荷、Open Food Facts 搜索与条码记餐、体重追踪，以及在本机读取体检报告。
一天内连发三个版本：v1.6.0（Tabata / EMOM / AMRAP 与更平缓的启动）、v1.6.1（离线训练、私密体检报告）、v1.6.2（照片又work了），窗口内共 30 次提交。
两条安全修复方向明确：fix(security): delete camera captures once read (lab report photos stayed on external storage)——读完即删相机拍摄文件，避免体检报告照片滞留在外部存储；fix(security): repair every untrusted shape in a backup——修复备份里每一个不可信结构。
最值得注意的是一份自曝的测试覆盖映射：docs: test-coverage map for 1.6.2 (95 rules: 53 pinned, 13 proposed, 21 manual, 8 unverified)——95 条规则被明确分成四档：53 条已钉住（pinned）、13 条是提案、21 条靠人工、8 条尚未验证。

值得关注的原因

「读完即删」这条正好治我化验报告 OCR 流程的一个隐患。 我的 GLP-1 营养伴侣规划了「化验报告照片 OCR → AI → App 导入」这条路，照片一旦落到共享存储目录，就等于把体检报告留在了不该留的地方。openfit 的做法是「读取完成即销毁原始照片」——这条应该直接写进我的导入流程规格。
「95 条规则分四档公开」是我见过最诚实的一份覆盖度声明，也是我专家包质检最该照抄的东西。 我 11-09 给自己的待办是「给规则加验证时间字段」，11-11 记的是「规则验证时间 + 失效扫描」，而 openfit 直接把规则状态分成「已钉住 / 只是提案 / 靠人工 / 尚未验证」四档并写进文档。 这比一个「验证时间」字段信息量大得多——它诚实地承认有 8 条规则连自己都没验证过。 这份四分法应该进我的规范文档模板。
它证明「自托管 + 本机 OCR + 无遥测」这套装配在真实项目里能跑通。 字体与 OCR 模型随应用一起分发，打开它或读一份体检报告都不会把内容发给 Google 或任何人——这正好对上我 4P 平台对数据主权的隐性要求（客户数据不出本地）。
对一人公司的现实意义：一个功能完整的自托管健康追踪应用，其工程量与复杂度已经接近我那 8 小时/周的极限。 这提醒我：「自托管健康 App」这条路我自己不该做，而应该是这类项目的用户与集成方。 我的价值在于精准营养的判断层与干预方案，而不是重写一个训练日志 App。

二、分类速览

① AI Coding / 编程智能体与技能生态

项目	星标	一句话
aequitas-labs/lobstah	★1	🦞✨ Harness-agnostic, token-efficient supervision framework for coding agents. Nobody stares at the water — your agents just bring home the lobstahs.
agbossjs/agboss	★2	Spring-style TypeScript framework for mission-critical LangGraph AI agents — @Application DI, Path Scope Flow DSL, transactional LIFO sagas, FinOps budgets, MCP/RAG/A2A & unit testing.
agentes-chat/agentes-chat	★260	Open-source, self-hosted Slack alternative where your team and AI agents work together in channels and DMs. Agents run on Claude, ChatGPT/Codex, Gemini or Grok, with MCP tools, their own computer and browser, routines, memory and approvals. Web, desktop and phone apps. Docker Compose + Postgres. AGPL-3.0.
agentrust-io/awesome-ai-governance	★66	🛡️ A curated list of tools, frameworks, standards, and resources for AI agent governance, safety, and compliance
AgentWorkforce/relay	★1,111	Infrastructure for coding agents
ah2233092/trading-agent	★0	Crypto futures watchman scanner + Claude Code deep-scan agent
AI4LGT/lamet-agent	★4	An Agent Framework for Large-Momentum Effective Theory Analysis
aliasunder/vault-cortex	★36	Standalone MCP server for Obsidian vaults — hybrid search, structured memory, task management, files, OAuth 2.1. Run locally, self-host, or one-click deploy for remote access.
alloverart/askhooman	★0	AskHooman: asks and tasks between people and their AI agents. An MCP server + app API, data in each user's own Google Drive
AmrDab/clawdcursor	★408	Local MCP server that gives any AI agent safe desktop control on Windows, macOS and Linux. Reads the screen as text and structure (accessibility + OCR) and acts by element name, with screenshots as the last resort. Multi-monitor aware, verifies every action, one safety gate.
Aquaticats/Monochromatic	★7	TypeScript and Rust monorepo: Claude Code and Pi coding-agent plugins, a minimal MCP stdio server framework, custom oxlint plugins and a pluggable Rust linter, policy-aware git tooling, core TS libraries, Wayland desktop apps (GTK4 file manager, Slint terminal, true-peak music player), OpenTofu infra automation, and a personal site.
araray/llmcore	★5	A unified, async Python framework for LLM applications—chat, autonomous agents, RAG, and sandboxed code execution. Supports OpenAI, Anthropic, Gemini, Ollama, and more through a single API. Features an 8-phase cognitive reasoning cycle, Docker-based isolation, human-in-the-loop approvals, session management, and comprehensive observability.
arisnachy/phoenix-harnes	★2	PHOENIX is a universal, provider-agnostic AI agent harness with intelligent model routing, persistent memory, tools, multi-agent orchestration, recovery, evaluation, and evolutionary self-improvement. Bring any API or run models locally.
ASTRANL/astranl-mcp	★0	AstraNL — protocol-governed coordination infrastructure for the AI + robot + human economy. Live MCP server, Europe-wide robot & drone database.
ayxuerui/contexture	★1	Context store for AI coding agents — markdown and git, catalog and graph retrieval, no vector database.
backblaze-labs/awesome-agent-infrastructure	★12	A curated list of AI agent infrastructure: memory stores, vector databases, execution sandboxes, MCP servers, tool registries, observability, and evaluation tools for building and running autonomous agents.
BaesTheorem/exobrain-harness	★1	A self-hosted AI exobrain. Claude Code plus 50+ skills, launchd watchers, and MCP servers, pulling voice transcripts, handwritten notes, calendar, tasks, and health data into one place so I stop losing track of things.
bilibiliUID1480494301/relay-hub	★2	Self-hosted LLM gateway — from LAN to the public internet (tunnel-friendly): multi-upstream key pool, downstream tokens, quota accounting, MCP & A2A routing, E2E envelope encryption. 自托管大模型中转站，局域网与公网两用，兼容 OpenAI/Anthropic。
Blaag/tfr	★3	Multi-world terminal client with typed JSONC configuration, a persistent multi-client gateway, LLM agents, and a plugin system with GitHub-based plugin installation.
busabase/busabase	★810	Open-source database & workspace for AI agents — structured data, durable knowledge, reusable skills, runnable apps, and human review on the writes that matter. Local-first and self-hostable.
calbebop/batesian	★3	Active security scanner for A2A and MCP servers
can4hou6joeng4/boss-harness	★2	面向招聘任务的 Agent Harness：可控工具、持久会话与桌面工作台
career-ops-hq/career-ops	★82,014	Open-source AI job search agent and job finder: scan job boards, score each job 1-5 against your CV before you apply, tailor an ATS-friendly resume and cover letter, get interview prep and an application tracker. It helps you fill in each application; you press Submit. Runs locally in Codex, OpenCode and other AI CLIs, or as a Claude Code plugin.
cellbang/cell	★748	The TypeScript framework that builds once and runs anywhere, designed for humans and coding agents alike.一次构建、随处运行的 TypeScript 应用框架，为开发者与 Coding Agent 共同设计。
chaklam-silpasuwanchai/research_skill	★1	My personal claude skill for doing research
chaterm/Chaterm	★3,107	Open source AI terminal for cloud and infrastructure management, enabling you to deploy, troubleshoot, and automate services using natural language and intelligent agents.
codecoradev/tole	★1	Durable Rust agent harness: chat-first personal assistant with mission mode — resumable after crashes, replayable forever
cosyte/agent-skills	★0	Agent Skills, Cursor rules, and AGENTS.md that teach AI coding agents when and how to use the MIT-licensed @cosyte healthcare data libraries: HL7 v2, MLLP, FHIR R4, X12, C-CDA, DICOM, NCPDP, and ASTM.
cporter202/agentic-ai-apis	★782	The ultimate collection of APIs for building autonomous AI agents — 2,036 production-ready APIs across Agents, AI Models, and MCP Servers. Stop wasting weeks building infrastructure. Plug these in and ship your agent today.
cuijiejie/open-agent-harness	★0	自托管的智能体平台：对话 + 工具 + 沙箱 + 按角色授权的 RBAC
D4Vinci/Scrapling	★101,681	🕷️ The BEST Web Scraping framework that handles everything from a single request to a full-scale crawl! Don't be shy, join here: https://discord.gg/EMgGbDceNQ and follow here for daily tips and tricks: https://x.com/Scrapling_dev
devbasex/ai-plugins	★1	AI Agent Marketplace - Claude Codeプラグインマーケットプレイス
developerz-ai/ultimate	★3	Bun-only, opinionated full-stack framework where the primary developer is an AI agent — and AI-maintained by developerz.ai. Eight primitives, one authz system across every surface, errors that carry an exact fix command. Start on a PaaS; scale out without changing app code.
Dicklesworthstone/agentic_coding_flywheel_setup	★1,676	Bootstraps a fresh Ubuntu VPS into a complete multi-agent AI development environment in 30 minutes: coding agents, session management, safety tools, and coordination infrastructure
dpc00/sublime-commander	★0	MCP server for Sublime Text 4 — lets AI agents read and control a running editor
dpc00/vscode-mcp	★0	An MCP server inside VS Code's extension host: run any command, edit files, install and test extensions
drptbl/logos-agents	★0	MCP server that lets AI agents hire each other over Logos, paid through an escrow, on top of logos-mcp
evolving-foundry/ember-rsi	★24	Make everything RSI-able! A pluggable framework for recursive self-improvement of AI agents and infrastructure through configurable strategies and evaluation-driven optimization.
facron2004/pi-desktop	★0	A Codex-style Electron desktop client for the Pi coding agent
fitzss/valun-open	★1	Valun for any AI agent, on any OS: MCP server, CLI and framework-neutral core for agent-to-agent service-for-credit deals; protocol spec with conformance vectors. Experimental pilot: synthetic units, nothing settled.
fr0ster/llm-agent	★12	Framework for building LLM agents, AI pipelines, multi-agent systems, MCP integrations, RAG workflows, planners, controllers and tool-driven orchestration across multiple LLM providers.
frankalcantara/Arestrix	★7	Arestrix is an AI Agent Operating System that provides a visual environment for building, coordinating, and running autonomous and multi-agent applications. It combines workflow orchestration, episodic memory, real-time observability, and LLM integration, allowing developers to design, test, and deploy intelligent agent ecosystems at scale.
FrozenYty/dsh-team	★0	Only the person changes, not the work. Retire, replace and recover agent-team members in DeepSeek Harness: the successor inherits the assignment, the session goes to a trash you control, and the freed slot comes back.
gekkyo/animate-mcp	★2	MCP server that lets AI agents read and edit the ActionScript/JSFL scripts in Adobe Animate documents.
gitpoo/imaginal	★0	What survives when an AI agent's memory is rebuilt: a synthetic-corpus harness, 45 saved runs and an invitation to break it.
gke-labs/kube-agents	★126	An autonomous agentic harness for Kubernetes. Proactive fleet audits, declarative GitOps remediation, and ChatOps in one place.
godic97/Gingoa	★0	An AI code guardrail and validation framework that strictly constrains and controls rogue AI coding agents and flawed codebases through logic and code.
gschaetz/humpyard	★0	Routing and policy gateway for LLM agent traffic: one Rust binary that speaks OpenAI Chat, Responses and Anthropic Messages, routes by request content, fails over, and records why.
hai-series/hai-framework	★4	Hai Framework is an AI-first, modular TypeScript full-stack framework that integrates LLMs, MCP, RAG, databases, authentication, security, type-safe APIs, Svelte UI, payments, and automated deployment, with built-in Agent Skills for AI-assisted development.
harshanandak/forge	★0	Forge — an open building-block platform for AI coding agents: blocks declare operations once; CLI and MCP are generated; works with any harness and any toolchain.
HoeTee/shipwright	★2	A plugin for Codex, Claude Code, and OpenCode that takes one coding task from plan to human merge.
homeassistant-ai/ha-mcp	★7,413	Configure and automate Home Assistant with Claude, ChatGPT, and other AI assistants.
IAI-Researching/Performance-Comparison-Demos	★0	Standardized autonomous benchmarks of frontier AI models across native developer harnesses (or dedicated agentic orchestrators) on demanding 3D physics applications with full performance telemetry.
IvanFlores/playwright-automation-framework	★0	Playwright + TypeScript automation framework with Claude Code skills (POM, test, spec and git standards)
JesunAhmadUshno/vllm-serving-lab	★0	LLM inference and serving, expanded in v3 to an inference optimisation lab: vLLM and SGLang, speculative decoding, KV-cache study, one benchmark harness.
jionpz/keel	★0	AI Engineering Runtime — 让 AI、人工、多 Agent、多模型、多 Harness 在同一套软件研发流程中长期协作的运行层
Jitsusama/agentic-harness.pi	★6	A pi package for collaborative coding workflows: planning, TDD, git hygiene, and code investigation.
K-kiron/dutyloop	★0	A repository super-agent for your coding tools. Plan, implement, verify and follow up with explicit budgets and human control.
kauki-labs/sandy	★1	Nix-based sandbox infrastructure and binary cache for the coding-agent microVMs
KAVentures/openworkgraph	★1	Open-source, local-first work context infrastructure for AI agents. Privacy-hardened workflow evidence via REST and MCP.
khaoss85/agent-crm	★2	Accordo — agent-native CRM framework for Claude Code, Codex and Gemini CLI. Author custom CRM and Customer Hub source with deterministic workflows, human approvals, audit and trace. Smart CRM is policy-governed; CDP + CRM adds bounded imports and logical identity, with no streaming ingestion, audience segmentation or CDP activation.
kingsoftcloud/agentkit-studio	★3	面向代码、资料与自动化任务的开放式 Agent 工作台：DSH 插件 + macOS/Windows 桌面端，覆盖 Agent 编排、资源管理、部署与自动化。An open Agent workbench for code, docs and automation — a DeepSeek Harness plugin with desktop builds.
koala73/worldmonitor	★90,497	Real-time global intelligence dashboard. AI-powered news aggregation, geopolitical monitoring, and infrastructure tracking in a unified situational awareness interface
KyaniteLabs/ag4te	★3	Git repository archaeology framework — mine commit history, detect signals, run 6 analysis vectors, and generate engineering reports. Python CLI + AI-agent ready.
langfuse/langfuse-docs	★253	🪢 Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.
Leoglez10/finanzas	★0	Local personal finance app for macOS, Windows and Linux in Rust: envelope budgets, scheduled payments, projected balance and an MCP server. Spanish UI.
liamphmurphy/vesper	★1	A frontier model chat-like experience, except memory, agent definitions, context, and more stay yours no matter the underlying harness you use.
lnardev/lean-agent-harness	★0	One rule file and one skill format for Claude Code, Pi, OpenCode, Codex, Cursor, Copilot and CommandCode. Standing context is AGENTS.md (about 0.9k tokens) plus one line per skill.
lobster0113/stalker	★0	Tournament poker AI study system, developed autonomously by an agent harness
localtaskrepo/lotar	★0	A git-native issue tracker that lives in your repo: tasks as YAML files, a local web UI, an MCP server, and coding agents that work your tasks in their own git worktrees.
luongnv89/herdr-github	★1	Open GitHub pull requests and issues of the current repo, in a herdr side pane
mchasse73/aegisx9	★0	Secure AI gateway: policy-enforced LLM access, prompt-injection and data-leak guards, audit, continuous red-teaming, GitOps delivery
mdiri-ai/SoduWork	★330	The AI-native engine for teams and their agents. Multi-harness, multi-agent, self-hosted.
michael-berardi/ultracompress	★3	Deterministic local compaction for Pi, Claude Code, Carl and Hermes Agent: VCC briefs, snap frames and lossless recall.
millathossain115/ai-forward-learning-client	★0	Interactive visual roadmap & architecture platform for AI engineers. Explore Algorithmic Foundations (DSA), Database Internals & B-Trees, Vector Embeddings, LLMs, RAG, MCP tool servers, and Autonomous Agents with live simulations. Built with Next.js 16, React 19, and Tailwind CSS. 100% client-side with zero backend dependencies.
minhtran3124/agent-harness	★1	Risk-gated /skills for Claude Code — 14 invocable skills carry a change from brainstorm → ship.
MoltyCel/moltrust-api	★4	MolTrust Protocol — AI Agent Trust Infrastructure
MrDHat/noma	★0	Run your existing coding agents in isolated, reproducible remote development environments. Bring your own harness, configuration, and infrastructure.
msaifkms/LLM-Gateway	★0	Production-style LLM API gateway with multi-provider routing, automatic failover, caching, rate limiting, cost tracking, and an evaluation harness.
MusadiqUrRahman/whaagent	★3	Multi-tenant WhatsApp AI agent platform — deploy intelligent agents on WhatsApp with LLM support, MCP tools, and production-ready infrastructure.
mvanhorn/last30days-skill	★67,350	AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary
openai/codex-security	★11,802	OpenAI's Codex Security CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities. npm: https://www.npmjs.com/package/@openai/codex-security
oscarka/deepseekharness-remotecontrol	★2	Control AI coding agents (DSH / Codex / Claude) from your phone — Feishu long-connection + self-hosted web console. 手机遥控 AI 编码任务：审批、插话、多机切换、E2EE 分享。No inbound ports.
Parcha-ai/agentrun	★460	AgentRun is an open-source platform that supercharges your agents. It adds self-improvement, durable execution and a control plane on top of your favorite agent harness. Your agents become smarter, more fault tolerant and use fewer tokens with every run.
philals/mission-control-capcom	★1	Mission control for story-driven work: local kanban boards, Claude Code skills and a live terminal board for tasks, draft PRs, CI and manual runs
pnedfff/claude-env-cleanup	★34	Codex skill for auditing and cleaning local Claude/CC Switch/coding-agent environment residue
polysius/nl-analytics-agent	★0	Text-to-SQL agent over Formula 1 data: LangGraph, hybrid retrieval in pgvector, read-only safety layers, Langfuse tracing, evals in CI, MCP server
puppyone-ai/puppyone-cloud	★1,299	The cloud infrastructure behind puppyone. Hosts projects as Git repositories for your team and agents.
Rayyan-Oumlil/OpenProxyAI	★2	Open-source LLM gateway: route by data class, audit trail, policy, budgets and MCP tool control.
reliant-labs/forge	★8	Go + React framework and CLI that builds, runs and ships static sites, SPAs, web/mobile apps and Go APIs (Connect RPC, protobuf, SQL migrations, generated ORM/handlers/hooks/tests), with linters for agent-written code. Deploys to hosted infra, Kubernetes, Docker Compose or host processes. Also an MCP server.
rgdevment/Tisty	★10	Local memory and context for you and your AI agent: Markdown notes, documents and tasks on your disk, with an MCP server for GitHub Copilot, Claude Code, Codex and Cursor. Open source, no account.
rlwillen0121/identity-eval-harness	★1	IdentityBench v4.1 & v4: multi-turn Enterprise IGA & Zero-Trust agent benchmark (Connectors, NHI Classification, NP-Hard Role/SoD Math, JIT/CAE, Access Reviews, JML) with multi-agent skills & steering. Published results run on v4_lite/v41_lite (60 scenarios, 360 tasks); harness supports v4_full/v41_full (180 scenarios, 1,080 tasks).
rnikadi/genaiopengateway	★0	Open-source GenAI gateway for every model type: LLMs, embeddings, text-to-speech, speech-to-text, realtime audio, image and video generation, plus an MCP and agent gateway. Enterprise features like SSO, RBAC, budgets, guardrails and audit are all included. Apache-2.0, with no paid tier.
roomi-fields/roomis-grillhouse	★1	An elephant–goldfish framework for building with Claude Code agents: a grill settles the architecture, fresh agents write tests, code and review each ticket.
sad-and-bad1231/awesome-LLM-inference-systems	★1	Awesome papers and open systems for efficient LLM inference, from foundational ideas to production serving.
SanvioLabs/sanvio-factory	★3	The smallest useful agent harness: instructions, a skill, a gate, a hook and a review agent. Works with Claude Code, Codex and Kiro.
searchware73/skills-maker	★0	Claude Code 自定义技能开发与发布：文献调研(lit-research) + 论文审查(paper-review)
shuruheel/mindgraph-mcp	★1	MCP server for MindGraph — connect Claude Desktop, Claude Code, and any MCP client to your knowledge graph
smithersai/smithers	★449	Smithers is an agentic workflow framework for defining workflows in simple TypeScript configuration files and executing them quickly, durably, and reliably
specnaut/specnaut-cli	★0	Specification-driven development framework — scaffolds a coordinated team of AI agents into your project, across every major coding harness
steelhead99x/rind-agent	★0	Agent skills for contributing to rind, the pure-Rust cryptography library behind MoQ
stevenmcconnon/multi-agent-finance-example	★0	A finance multi-agent system, end to end: SEC filings ingested into a vector database, permission-filtered retrieval, a supervisor graph coordinating deterministic and model-driven agents, human approval, audit and replay, a web app, and AWS infrastructure. A reference architecture for running AI agents on private documents.
superserve-ai/superserve	★507	Sandbox infrastructure for AI Agents
tenalisriharsha/shadowci	★0	An ambient coding agent that watches your work, fixes bugs in a shadow git worktree, and shows a verified patch with proof only when sure. Includes a replay-based evaluation harness for detection latency and false interruptions.
TerrenoLabs/terreno	★3	Batteries-included full-stack TypeScript framework built for AI coding agents: auto-generated REST APIs, auth, admin, and local-first sync, on one React Native app for iOS, Android, and web.
TimothyVang/deep-mode	★0	Open-source AI code review and software factory for coding agents: ACP agents (omp, Claude Code, Codex) review, fix, verify and merge your PRs and build your issues, on your own subscription.
tobocop2/lilbee	★70	A local model engine with a search engine and web crawler built in. Run your coding agents on your own GPUs, and get cited answers from your files, code, and the sites you crawl. One executable.
Toolbox-POD/toolbox-pod-mcp	★0	Toolbox POD's hosted MCP server: design custom shirts and hoodies, see mockups, and order prints from your AI app.
true-line/agent-bot-identity	★0	Give every coding agent on your machine its own GitHub identity — per-harness GitHub App bots, git-native hooks, and a fail-closed gh.
TruffleCraft/feedbackkit	★1	Self-hosted feedback gateway — complete-at-the-source user feedback → structured, agent-ready GitHub issues. Privacy-first, no session replay, local-LLM-ready.
TTNAN/quark-skill	★2	一句话把网盘资源存进夸克的 Agent Skill：PanSou 搜索 + 夸克官方转存，支持 Claude Code、Codex、Cursor 等。
twinnydotdev/twinny	★3,695	Open-source AI coding assistant for VS Code. Coding agent, code completion, chat, edits and reviews with local or hosted models. Your models, your infrastructure.
use-junction/usejunction	★11	AI coding spend management for engineering teams. Track Cursor, Claude Code, Codex, Copilot, usage, cost, plan utilization, seat waste, and device health.
vanillagreencom/kendex	★93	Package manager for agents, skills, hooks, and extensions. Author once, install on every harness. QOL features included.
viniciusmachadovianna/skill-shard	★0	Codebase and evaluation framework for my Bachelor's Thesis (TCC), focused on developing a Relational Retrieval-Augmented Generation (RAG) system that combines vector search, cross-document entity linking, and explicit source citation to minimize LLM hallucinations
weideyouni/holoaeon	★0	A Zero-Dependency, Bio-Inspired Cognitive Framework for General-Purpose Multi-Agent Systems
whichguy/skill-craft	★1	Host-neutral portable skills monorepo (skill-interop and more). Not claude-craft.
zch6a/build-your-own-ai	★0	Build your own AI from scratch: tokenizer, autograd, GPT, inference, vector database, RAG, agent and MCP. Each a small, tested, readable module that runs on a laptop CPU with no API key.
zchoi/Awesome-Embodied-Robotics-and-Agent	★1,916	This is a curated list of "Embodied AI or robot with Large Language Models" research. Watch this repository for the latest updates! 🔥

② 医学 / 临床研究 AI 与 Agent 技能库

项目	星标	一句话
1ali2003ah1-lgtm/ProtonAI	★0	Clinical AI Platform for Proton Therapy
agustinroviraquezada/clinical-deid-es	★0	Master's Thesis for the Master's in AI Development (BIG school): A web-based tool that removes personal data from a clinical note in Spanish before using it with a third-party LLM. A proprietary model (a clinical RoBERTa fine-tuned with MEDDOCAN) detects personal data; only questionable cases are submitted to an LLM, and only the affected phrase is
Ahmed28A/ovarian-cancer-dual-stream	★0	Interpretable dual-stream AI for ovarian cancer: ConvNeXt+CBAM+Grad-CAM subtype classification (99.29%) and CatBoost+SHAP clinical prediction (AUC 99.35%), with a Gradio demo.
alexblogoodf/glyfiq-mcp	★0	Hand-drawn medical & health icons for AI agents (MCP server + Claude skill)
alvianzf/AE3	★0	An AI clinical assistant that only answers from your practice's own curated library — every claim cited, every answer independently checked before a practitioner sees it.
andreibesleaga/AgenticMedicalImagingHelper	★2	Agentic AI analysis of medical image series over time. Research and education only, not a medical device [Demo]
AndrewPJ2002/AI-Clinical-Mental-Health-Support-Platform	★1	A platform that connects patients and clinicians on the basis of mental health conversations
Annapurnar06/alterra	★0	Healthcare AI Safety & Alignment Platform — simulation environments, evaluation, post-training, and RL infrastructure for clinical AI
assiaezzayani-developer/GreenShield-AI	★0	GreenShield AI is an intelligent plant health platform powered by AI, delivering advanced plant disease diagnosis, personalized care recommendations, health tracking, and recovery monitoring. Designed to help individuals and growers protect plants, improve plant health, and promote sustainable agriculture worldwide.
awslabs/hcls-agent-skills	★76	Open-source healthcare AI agent skills and life sciences agent skills. 42 skills for medical coding, risk adjustment, claims analytics, quality measures, genomics, imaging, and drug discovery across 13 HCLS domains.
berntpopp/litvar-link	★0	MCP server for NCBI LitVar2: search the biomedical literature by genetic variant — the index linking variants to the publications that mention them, as typed tools for LLM agents.
CaiZongyuan/clinmesh	★3	面向 Agent 评测、产品验证和医疗信息系统研究的中国公立医院 HIS 虚拟环境，提供 FHIR R5 接口与可重复、可审计的合成业务场景。
Chetanprince213/healthcare-ai-data-platform	★0	Healthcare Data Engineering, Analytics, MLOps and Agentic AI Portfolio Project
concept2cure/ClinicalSageAI-2-replit	★0	Repository for https://replit.com/t/clinicalsage/repls/ClinicalSageAI-2
damsolanke/aarogya-atlas	★2	Agentic healthcare-facility search for India: a tool-calling loop over FHIR-aligned data with a critic pass on every answer. Hack-Nation 2026 / Challenge 3 (Databricks).
dareinelt/lanpa	★0	Comprehensive intranet solution for healthcare institutions, integrating a proprietary chat agent, Nextcloud, EuroOffice, as well as an alternative web-based Exchange client.
dhruvsave29/cognistream	★0	CogniStream - Winner, COHERENCE Hackathon 2026. AI clinical-trial matcher: 30+ min screening to under 15 seconds (RAG + Groq LLM, FastAPI, Next.js).
elong0527/ai4csr	★3	AI for Clinical Study Workflow
farruxshernazarov5-dotcom/med1.uz	★0	MED1.uz is an advanced medical ecosystem integrating 72 AI agents and 14 AI models for early diagnosis of oncology and diabetes, as well as clinic digitalization.
Gaogxx/MedicalGPT_agent	★0	A medical large model agent based on MedicalGPT and DeepSeek Harness, covering model training, capability alignment, and runtime framework development and deployment.
gil906/devpost-buildhybridai-2026-31808	★0	PII-Zero: Zero-Trust Clinical & Legal Hybrid AI Airlock (#BuildHybridAI 2026 — Web + Chrome MV3 Extension + Kotlin Compose Foldable Android)
godasrinivasulureddy/Fame-Genix-Agentic-AI-2.0	★0	Multimodal, agentic AI-powered health monitoring, risk assessment, medical document understanding, RAG-based health guidance, and coordinated emergency response.
Guitrentz/clinical-ai-deterministic-harness	★0	Deterministic clinical validation harness in Python for AI-augmented healthcare systems: SNOMED CT (SCTID) Verhoeff validator, US NPI/DEA/FDA UDI & Brazil CNS checksums, and HIPAA Safe Harbor PHI scrubber.
HusseinYassineMD/ClinicalCore_Aim1	★0	USC ADRC Clinical Core Aim 1 metrics dashboard
ilay-fisher/vitalstream	★0	Real-time clinical telemetry platform with an AI clinical copilot: Kafka, Spark, Spring Boot, Cassandra, Elasticsearch, Redis, RabbitMQ, pgvector RAG, MCP, Kubernetes
JoshuaOforisurg/surgical-data-platform	★0	data engineering platform for operating theatres, supporting operational pipelines, analytics, and AI-ready datasets from fragmented clinical systems.
KennethJSmithDev/DICOM3D_AI	★1	DICOM3D AI Hub uses a native InterSystems AI Hub agent in an IRIS Interoperability workflow to validate FHIR imaging evidence and select source-linked findings in an interactive 3D anatomy viewer.
kollaprudvi79-ai/clinical-trial-operations-dashboard	★0	Portfolio-grade Clinical Trial Operations Dashboard in Power BI — synthetic data, star schema, exact DAX, 5 report pages
lakshmisanthosh123/healthcare-ai-platform	★0	Enterprise AI Healthcare Platform built with React, Micro Frontend, FastAPI, PostgreSQL, RAG, AI Agents, LangGraph, Docker, Kubernetes, and AWS.
lawrencelucado/coding-audit-copilot	★0	AI medical coding agent: Claude + deterministic CMS-rule tools (ICD-10-CM, NCCI, MUE), scored against CPC-reviewed answer keys, with cost/latency tracing and a CI evaluation gate.
mbzuai-ai4bio/BioXArena-Pro	★1	BioXArena-Pro: a research-grade benchmark for autonomous biomedical machine learning agents (harness, graders, scoring, leaderboard)
mianaz/biomed-skills	★0	Claude Code / Codex agent skills for biomedical research: R tidyplots and GraphPad Prism figures, paper framing and writing, protocols, Methods and mechanism animations.
mohammadazam7/clinical-data-lakehouse	★0	FHIR R4 patient records to a tested OMOP CDM 5.4 warehouse on DuckDB and BigQuery, with de-identification, data quality checks and care gap measures. Synthetic patients only.
muhammadusamakhan911-bit/fabric-mlv-agent-healthcare-starter	★149	Fabric MLV healthcare demo template: lake views, ontology, data agents
pedropcamellon/folium	★0	Folium is an engineering sandbox for agentic AI in healthcare — not a commercial or feature-complete EHR.
Prashanthbkm/MediLens-AI	★1	🩺 MediLens AI – An AI-powered medical report analyzer that extracts health data from PDF reports, explains test results, tracks health trends 📊, and provides intelligent insights using RAG 🤖.
pv-tech28/PCOS360	★0	PCOS360 – AI-powered PCOS Health Intelligence Platform \	Track symptoms, discover personal health patterns, understand lab reports, and gain meaningful health insights.
Rhushabh1/SugenICD	★0	AI-powered Clinical Coding Engine
saikumarreddy2902/MediSphereAI	★0	AI Agent that helps in Healthcare.
sisovin/healthcare-ai-desk	★0	M1.1 is infrastructure scaffolding only. Authentication, organization/project modules, database migrations, agent orchestration, and clinical workflows are not implemented yet. No tests are claimed to have passed. Run the verification steps below on your computer.
smaniches/uniprot-mcp	★3	Auditable UniProt MCP server with per-response SHA-256 provenance, release pinning, verification, and offline replay.
Snigdha-0210/Cancer-multi-agent	★0	Autonomous, state-driven multi-agent oncology intelligence system powered by LangGraph, 103-document biomedical RAG (31,498 units), local Ollama LLMs, deterministic safety guardrails, and fail-closed clinical verification.
suresh121011/SEHAT-AI	★0	SEHAT AI is a multimodal healthcare triage assistant designed for Indian healthcare facilities. It combines voice, text, and medical report extraction with rules-based urgency assessment, evidence-linked summaries, and human review to support safer, faster clinical handoffs. Built with a focus on accessibility, privacy, and offline resilience.
Swethavenna-github/clearbill	★0	Clearbill: two AI agents settle a medical bill. PayPal + AI Hackathon prototype (fictional data, PayPal sandbox).
veganbeef-96/lastehr-platform	★0	Open-source platform for building AI agents on FHIR health records, including the agent app, MCP server, and conformance tools.
VenkatesaPerumal-23/HealthCare_Agent	★0	An agentic healthcare information assistant that answers questions using local PDF references and trusted health websites. It is intended for educational use only and is not a substitute for professional medical advice, diagnosis, or treatment.
wetware-hq/chart	★0	An agentic medical chart.
xuyul6/med-agent-runtime-v1	★1	Medical information multi-agent runtime rebuilt from requirements, with source-grounded answers and observable execution.
zlatedod/multi-agent-system-for-healthcare	★0	MAS-mcp healthcare research: n8n workflow, agent prompts, cohort SQL, and setup instructions.

③ 健康管理 / 营养与食品 AI

项目	星标	一句话
12alankrita11srivastava2610-droid/nutrition-agent	★1	AI Nutrition Assistant powered by IBM Granite
A0966411725-png/dietsense-proposal	★0	DietSense：對話式每日熱量攝取 Domain AI Agent — 以 OpenSpec 重構的期末專題提案
aderaaij/loopback-training-server	★9	Self-hosted backend for Loopback – Run Coach: workout tracking, Apple Watch training queue, training plans, health metrics, web dashboard, and an MCP server for AI coaching
Amna-Coder404/ByteDiet	★0	ByteDiet is an AI-powered personal nutrition tracker built with React Native and Expo. Analyze meals, estimate calories, track daily nutrients, and get personalized food insights using Google's Gemini API.
ayhamimarati/Bourmdane-mohamed	★0	AI-powered nutrition and calorie tracking app
bigidrise/my-perfect-meals-npm-dev	★0	My Perfect Meals is an AI-powered nutrition and meal-planning platform that personalizes meals around individual health needs, food preferences, and lifestyles. Copy
chakravartiavinit/relife	★1	A second chance at life: local-first tracker for health, mind, skills, wealth and relationships. Telegram reminders, Binance/GitHub/bank-SMS connectors, bring-your-own-key AI.
DavidRamms/plant-based-research-hub	★0	Daily-updated plant-based diet research tracker with AI-generated summaries
DragonScripter/pulse-health-analytics	★0	A health tracking platform featuring a native Kotlin Android client, a Node.js API gateway, and a Python AI engine for biometric trend analysis(Not completed)
drew-weise/Provisions	★0	AI-assisted personal food planning for meals, groceries, budgets, nutrition, and preferences.
Dzarlax-AI/health_dashboard	★1	Personal health dashboard — ingests Apple Health data via Health Auto Export, tracks sleep, HR, HRV, SpO2, steps and readiness scores. Web UI + MCP server for AI assistants
edisontaimu9-ui/thanzi-screening-engine	★0	AI-assisted malnutrition screening engine with deterministic WHO anthropometric calculations, ML support, clinical classification and explainability.
ehtesham8958873-ops/caloriewise-ai	★0	Free online calorie calculators and nutrition tools to help you track daily calories, BMR, TDEE and macros.
faisal-abdelnase/NutriVision	★1	NutriVision is an AI-powered dietary health assistant built with Flutter that helps users track meals, analyze nutritional data, monitor dietary goals, and receive personalized AI-driven nutrition insights, healthy alternatives, and recipe recommendations.
ganeshvarun40-maker/BMI-Health-Dashboard	★0	A BMI Health Dashboard built with Python and Streamlit, featuring an AI Nutrition Coach, diet plans, charts, and health tracking.
git-macko/resolution-fitness-app	★1	A full-stack gym fitness mobile app -> Go REST API backend with SQLite + React Native (Expo) frontend. Plan routines, execute workouts, scan food for nutrition insights, track progress with XP/streaks, and chat with an AI fitness coach.
Javiguegadev/personal-trainer	★0	AI-powered nutrition & fitness app (freemium). Web-first Turborepo monorepo: Next.js, Hono, Supabase, Drizzle, Claude API — mobile-ready with Expo.
jollyjoker992/garmin-archive	★0	Back up your Garmin Connect data (official Personal API) to Google Drive and S3 Glacier, map it to FHIR R4 and load it into a local Medplum. Docker, daily scheduler, MCP/GraphQL for agents.
Justusability/license-label	★0	A nutrition label for the open-source licences inside your app. Plain language, for people who build with AI coding agents.
kdinya/smart-diet-ai	★0	A simple food diary with AI-estimated calories and nutrition, daily balance and weight tracking. English and Ukrainian.
lgf2111/snapbite	★0	Telegram-native AI calorie tracker — send a photo or describe a meal, get an editable nutrition estimate. Cloudflare Worker + Telegram Mini App, BYOK.
manoj234-pull/DASH-Meal-Prep-AI	★0	An agentic DASH meal-prep assistant that finds recipes from available ingredients using DASH diet.
merickvaughn/lifting-logbook	★0	TypeScript training log with AI-driven recommendations based on performance, sleep, and nutrition
mkpoli/health-tracker	★1	AI-assisted multilingual health record tracker for lab results and clinical metrics
munimtechnologies/open-food-safety-check	★0	Pre-registered checks that published food-safety survey numbers re-compute from the survey's own open data. By MV2 of Munim, Inc., made with AI agents. Apache-2.0.
muscleshubfit-cpu/alkemos-aws	★0	Alkemos (alkemos.com) — production repo of the bilingual AR/EN smart fitness & nutrition platform: EVO AI coach, exercise & food libraries, planning tools, B2B coach system, memberships & full admin — Next.js 16 · Supabase · AWS Lightsail + Cloudflare
nikhilramakrishna3005/nutricart-ai	★1	The AI for food nutrition
NurtureAISpace/NurtureAISpace.github.io	★0	NurtureAISpace: Autonomous Space Nutrition & Bio-Regenerative Life Support System (N-BLSS) for Mars Sol-450 Missions — Thermodynamic Arrhenius Kinetics, SciPy NASA-STD-3001 Diet Optimization & Habitat Simulation Engine (FastAPI + SciPy + Python)
OSerHuang/caldoo	★0	卡嘟 (CalDoo) - offline-first Android diet & nutrition tracker (BYOK AI)
PatricioMR/nMRPro-H---1D-Metabolomics-Platform	★0	Transparent, auditable pipeline for 1D 1H NMR metabolomics: signal detection, multiplet-based annotation, quantification with detection limits, statistics and PDF reporting.
PedroPinho17/calorie-tracker	★0	Snap a photo of your meal and get calories and macros: AI food recognition (Gemini) matched to INSA, USDA and Open Food Facts nutrition data. NestJS API, Next.js web and Expo mobile app, PostgreSQL.
petrovski-stefan/nutritionals	★0	North Macedonia's first ever supplements/OTC web aggregator for price comparison, AI-powered search included
PrimeCaster/JustGetFit	★1	All-in-one Android fitness, strength training, cardio, and precision AI nutrition macro tracking app built natively with Jetpack Compose and Material 3
prokesmic/HealthUniverse	★0	Living knowledge graph of nutrition/lifestyle/supplements vs chronic disease — Claude-seeded, Gemma-maintained
raslmm/safebite	★1	AI Agent that detects allergens from food labels
Rmjjke/BabyBloom	★0	iOS app for newborn health tracking — feeding, sleep, diapers, growth with AI assistant
RuneYang123/fitness-companion	★2	Fitness Companion \	A Chinese-language fitness skill for adults, tailoring training, nutrition, supplement choices, and sleep guidance to your goals, schedule, and equipment. Supports exercise tutorials, on-demand feedback, and portable plans for continuing across chats.
rupal2k/openfit	★1	Self-hosted training, body-weight and nutrition tracker with on-device AI vision: guided workouts, progressive overload, passkey sync, PWA and standalone Android.
sabasgh/food-rag-chatbot	★0	Three food search systems on one ChromaDB vector store: an interactive CLI search, advanced filtering by cuisine and calories, and a RAG chatbot powered by IBM Granite. Built during the IBM RAG and Agentic AI Professional Certificate.
Sculpt-AI/nutrition	★0	Calorie and macro engine behind Sculpt AI's app, server and website
Smith3939/smit-gym	★1	Smit Gym - Fitness & Nutrition App with AI Coach \	React Native + Firebase + Claude AI
Soorya2201/Immuny	★0	Clinical AI allergy tracker combining a vision-language model (MedGemma), IoT wearable monitoring, and voice UI for proactive allergen management.
spoonjoy/spoonjoy-apple	★0	The native SwiftUI app, with a cook mode that walks you through a recipe one step at a time.
ulko31/sofe	★0	Wellness Telegram Mini App: nutrition, workouts, progress tracking, social features, maps and AI assistant
xLightBringerx/pocket-coach	★1	An AI driven fitness app that can create weekly and monthly fitness regimens and diet plans and book clients with personal trainers in their area.

④ 模型底座与 Agent 基础设施

项目	星标	一句话
ai-verification-observatory/inference-proofs	★10	Provable LLM inference measured on the same hardware: overhead against production serving, proving time, proof size and check time, and what one proof establishes.
aidless/kv-serve-lab	★0	CPU-verifiable toolkit for LLM inference serving: KV cache sizing (GQA/MLA), PagedAttention block pool, continuous batching simulator, and SSE streaming bench. Stdlib only, 180 tests.
aiverse-filmmakers/AI-Verse-Memory	★0	Local-first persistent memory for AI agents, with no external database, vector store or extra API required.
andrewyng/openworker	★19,006
AndryanZulfi/hermes-server-stack	★1	Hermes Agent Production Server Infrastructure Stack (9router, Honcho, Monitoring, Galaxy View)
apache/pinot	★6,171	Apache Pinot: a real-time analytics database for low-latency SQL and vector search, serving user-facing and agent-facing applications
api7/aisix	★219	Open-source AI gateway for LLMs & AI agents, built in Rust. One OpenAI-compatible API for OpenAI, Anthropic, Gemini, Bedrock & more — routing, guardrails, caching, rate limits, observability.
awslabs/unified-kg-rag-on-aws	★39	AWS-native knowledge graph RAG framework. Runs GraphRAG, LightRAG, HippoRAG 2 and RAPTOR on one shared index, each selectable at index and query time, with hybrid BM25 + vector + graph retrieval and source provenance. Amazon Bedrock, Neptune and OpenSearch, or fully local. Incremental indexing, multilingual, bring-your-own KG, built-in evaluation.
beehive-lab/jitllm	★295	Java Inference Tornado toolkit: High performance LLM inference & serving for the JVM on GPUs
bowenliang123/dsh-context	★4,961	The best DeepSeek Harness plugin for context insight and management, with context dashboard / browser / sidebar and context command, for context statistics, composition, breakdown, evolution details, agent teams, understanding how the context is made of, and how it evolves. 一站式 DeepSeek Harness 上下文可视化插件，Context 面板、浏览器、侧栏、命令，透视上下文组成、演进、压缩、多 Agent等。
cookerpapa/pi-managed-agents	★8	Self-hosted infrastructure for managed agents, powered by Pi.
cosyte/cli	★0	The Cosyte command line for Node.js and TypeScript: one command over HL7 v2, FHIR R4, X12, ASTM, NCPDP, C-CDA, DICOM, and MLLP, with HL7 v2 to FHIR conversion from transform, code mapping from terminology, and the cosyte-mcp MCP server for agents.
dakotacookson/Llama-RPC-Farm	★8	Web UI to pool your home machines into one llama.cpp cluster: multi-machine RPC GPU pooling + on-demand multi-model serving via llama-swap
Drgr33nSenior/Spry.ai-rdna4-serving	★1	Experimental LLM serving on AMD RDNA4 GPUs, with ROCm/SGLang integration, FP8 model support, four-bit KV caching, and reproducible benchmarking.
elnomic/fesnuk	★1	Open-source programmable wallet infrastructure for AI Agents - Golang-first.
evoila/meho	★22	Governance backplane for AI agents acting on infrastructure — policy-gated, audit-grade, MCP-native. Apache 2.0.
farming-labs/farm.js	★659	a full-stack, renderer-agnostic framework for the web and agents, with an experimental AOT React compiler.
fluffyled/enterprise-rag-framework	★0	Local RAG framework for employee navigation through internal regulations: hybrid search, OCR and grounded citations.
ginuerzh/cuttle	★1	Cuttle is a self-hosted LLM gateway with an embedded admin console.
honeybee-studio/honeybeepf-llm	★11	Lightweight eBPF-based observability agent for AI/LLM workloads - monitors token usage, latency, and sensitive file access in Kubernetes without code changes.
humemai/tiervec	★0	Tiered vector index for agent-memory databases on the JVM: exact tail, sealed graph segments, scalar-safe kernels. Apache-2.0.
hyunnnchoi/vllm	★0	A high-throughput and memory-efficient inference and serving engine for LLMs
jingwenzz/miniserve	★1	A minimal LLM inference serving simulator with continuous batching (Python)
joseoper/token-diet	★0	Audit token footprint, detect prompt bloat, and neutralize silent context leaks in AI agent workflows
junxnpng/junxnpng.github.io	★0	Personal site — paper notes (Korean) on LLM inference & serving, built with Hugo + PaperMod
khaneilcampbell/ai-ops-atlas	★1	Atlas is a secure, observable AI operations control plane that unifies enterprise infrastructure, repositories, business workflows, and autonomous agents into one intelligent operating platform.
luharry2821/pingo-voice-backend	★0	`Pingo voice tutor backend: real-time language learning agent built on AssemblyAI Universal-3.6 Pro Realtime STT, LLM Gateway, and TTS.
luqman-v1/9router-go	★282	High-performance Go proxy gateway for 9Router LLM routing — 32K RPS, 42MB RAM, CGO-free
MacCracken/hoosh	★4	AI inference Gateway — multi-provider LLM routing, local model serving, speech-to-text, and token budget management
masterledgerlive/koda-kai-decoder	★0	Koda Kai the Decoder — the open universal decoder: every cipher and codec in library form, container interrogation of untrusted code, and public agent report cards. MIT licensed, free for everyone.
mfittko/dev-loops	★18	Shared AI agent workflow infrastructure for reusable local and remote development loops
mihazs/agent-harness-lab	★4	Evidence-driven agent harness diagnosis, process mining, and paired evaluation
mrveiss/AutoBot-AI	★40	Self-hosted, agentic AI you own — voice, vision-driven browser & computer control, visual workflows, human-in-the-loop approvals, multi-user RBAC, and a knowledge graph. Runs on your infrastructure. Your data. Your AI.
nfsnyy/signet	★0	Acceptance infrastructure for human–agent development: specs frozen before execution, machine-checked gates, every claim auditable. CLI-first
nikoma/bicdb	★3	Next-gen database for AI agents. PostgreSQL, Redis, queues, vectors, full-text search & analytics. One engine.
npspringmeyer/clinical-guideline-rag-eval	★0	A RAG assistant that answers questions over public health guidelines, with a measurable evaluation framework
olathedev/hedgerow	★0	A privacy router for the Otari LLM gateway: sensitive prompts stay on a local open-weight model, everything else goes to the cloud.
openai/openai-agents-python	★30,623	A lightweight, powerful framework for multi-agent workflows
pingcap/tidb	★40,732	TiDB is built for agentic workloads that grow unpredictably, with ACID guarantees and native support for transactions, analytics, and vector search. No data silos. No noisy neighbors. No infrastructure ceiling.
PNUops/pickle-llm-gateway	★0	LLM API 게이트웨이 (Go)
Praneeth16/kumo-fraud-databricks	★0	Fine-tune NVIDIA Kumo-Tabular on Databricks AI Runtime and serve it on GPU Model Serving with Lakebase features: card-fraud scoring in real time and in batch
pratiksinghchauhan/framework-compass	★0	A RAG assistant for NIST AI RMF and OWASP LLM Top 10, with a public evaluation report.
radiatedsaturn7/automated-llm-proxy	★0	automated llm gateway
rafiebazine/rag-image-captioning	★0	Contextual Compression RAG framework for automatic image captioning on MS COCO. Grounds Vision-Language Models (VLMs) by dynamically compressing retrieved multimodal context to reduce redundancy, lower object hallucinations, and improve caption precision.
railgrid/railgrid	★14	Railgrid helps platform teams turn company data, tools, and infrastructure into reusable building blocks for AI-built apps and agents.
ridsood90/agentic-ai-lab	★0	Building agentic AI systems with Python, LLMs, tool calling, memory and observability.
riffdb/riffdb	★0	RiffDB is an ACID-compliant relational database for agent-built SaaS that unifies transactional writes, columnar OLAP analytics, and full-text and vector search in one contract-first engine.
ROCm/rocm-cli	★51	ROCm CLI is a command-line tool for setting up and running local AI on AMD GPUs, with a full-screen TUI dashboard for GPU telemetry, model serving, and chat.
rrnewton/agent-utils	★3	A simple infrastructure for running a DAG of internally-parallel steps for CI with profiling planning and resource limitting.
sankalp-rajeev/Nexus	★0	LLM observability and self-healing platform for detecting, diagnosing, and repairing failures in multi-agent AI systems.
seckinozbek1/project-shimmer-deployment	★0	The deployment-phase version of Project Shimmer, an agentic framework to verify the claims written in the document and detect the changes in political, legal, or diplomatic documents between negotiation rounds.
Senguk520/dsh-multitask	★52	A Multitask "shock absorber" plugin & agent preset for DeepSeek Harness (DSH): the main chat only plans, briefs subagents and summarizes results, while file reads, commands, search and builds all run inside subagent contexts. Four worker types, an approval gate, zero DSH source changes.
Spongy01/Conduit	★0	LLM Gateway with Rate Limiting, Fallback Routing, and Observability
storytold/craft-anything	★28,556	Craft Anything: understand how any software works, and craft your own.
sudarshan1x1/efficient-inference	★0	Benchmarking LLM serving under concurrency on a T4: 22.6× throughput from batching, plus where KV cache, prefill and quantization help or don’t.
supportauranova23-max/-optimax-llm-router.	★0	"An ultra-fast, data-saving LLMOps routing gateway that intercepts AI prompts, applies dynamic semantic caching, and optimizes API compute workloads to drop enterprise model costs by up to 60%."
tachyon-zcash/ragu	★61	Rust language proof-carrying data framework
Theorvane/OpenGranter	★0	IAM-aware LLM gateway with OpenRouter and direct provider routing
westhong/astro-dashboard	★0	A personal dashboard by K3 to review my astro shooting conditions for my desired locations
xiseliuli/dsh-as-mcp	★6	Expose a running DeepSeek Harness as an MCP server, so Claude Code, Codex or any MCP client can drive its workspaces, sessions, files, shell and agent tools, and read token usage and cost receipts.

三、今日趋势观察

⭐⭐⭐ 技能（skill）正在从「提示词片段」变成「可携带、可评测、可更新的方法论资产」，并且已经有了公开的工业级样本。 本期 awslabs/hcls-agent-skills 把 43 个医疗与生命科学技能按 reasoning（方法论与判定框架）与 pipeline（工具链与参数表）二分，配一份 410 题两两对比评测证明技能版胜出 70~86%。这与我连续四期记下的主线完全对上：11-08「测试失败要报具名场景」、11-09「规范/文档/矩阵对齐代码 + 回归守卫」、11-10「让不知情的第三方来判」、11-11「读声明必须配文件清单」——这一期这些机制第一次被装进了一份可以下载的资产清单。
⭐⭐⭐ 证据链开始从「结论」向两端延伸：一端是修改的来源与归属，一端是实体重命名后的历史结论。 codex-security 把 AWS 生成的补丁带进第三方扫描器（ArmorCode / AWS Continuum）的发现条目，并导出 SARIF/JSON/CSV；unified-kg-rag-on-aws 则把「重命名」做成一次需要幂等键、需要保住被引顶点的迁移。合起来看，成熟系统正在回答同一个问题：三个月后我怎么知道当时那份报告、那次修改、那个术语是怎么来的。 这正是我 GRPM 报告可信度三段路径（数据归档 → 判定规则 → 可复现制品）尚未走完的部分。
⭐⭐⭐ 健康线的分水岭从「能不能记」变成了「数据留不留、规则诚不诚实」。 openfit 把体检报告 OCR 做成了「读完即删相机照片」，并公开自己 95 条规则的验证状态（53 已钉住 / 13 只是提案 / 21 靠人工 / 8 尚未验证）；uniprot-mcp 用每次响应的 SHA-256 溯源、版本钉定与离线重放把蛋白质数据做成了可审计件；openmed 与 SAGEAgent 一个给 72 个医疗技能装上本地脱敏流水线，一个让 agent 判断「下一项检查值不值得做」（平均检查负担降 55% 而 C-index 不显著下降）。共同点：这一期的健康项目都在回答「我的数据去哪了、我的结论凭什么可信」，而不是「我又加了个拍照算卡路里的功能」。
⭐⭐ 架构层的边界比提示词层的约束更硬。 clinmesh 明确写着「最终 Actor 始终是点击原生审阅框的登录人类」——Agent 拿不到提交按钮；codex-security 把「配置错误」与「策略拒绝」拆成两类诊断；unified-kg-rag-on-aws 默认用无 LLM 的三项指标、LLM 裁判只是可选项。这三条都在说同一件事：把约束从「请不要这样做」变成「系统不允许这样做」。
⭐⭐ 授权形态的三档分流在本期依然成立，而且有一条例外值得单独记住。 本期 273 个条目中，宽松型授权（MIT / Apache-2.0 / BSD / MPL / CC0）共 174 个；传染性 copyleft（AGPL / GPL / LGPL）11 个，属源码可见非开源的 1 个，专有授权 1 个；另有 86 个条目根本没有 LICENSE 文件——方法论可以读、代码不能直接依赖，这条边界每一期都要重新确认一遍。

四、GitHub 地址速查

GitHub AI 日报 · 2026-11-12（周四）

一、今日精选 5 条

1. awslabs/hcls-agent-skills：AWS 把 43 个医疗与生命科学技能做成了「可移植的标准件」，并用 410 道两两对比题证明它比不加载技能更强

2. openai/codex-security：官方给 agent 装上了「安全扫描器」，并且开始把别的扫描器的结果也纳入统一证据链

3. awslabs/unified-kg-rag-on-aws：四种图谱 RAG 跑在同一套索引上，并且开始把「重命名」当成一次需要回滚的迁移

4. CaiZongyuan/clinmesh：把中国公立医院的挂号—医嘱—药房—医保—病案做成可评测的虚拟环境，而且明确「最后落笔的必须是人」

5. rupal2k/openfit：一天发三个版本，把自托管健身 App 的体检报告 OCR 做成了「读完即删」的隐私边界，并公开了自己 95 条规则的验证状态

二、分类速览

① AI Coding / 编程智能体与技能生态

② 医学 / 临床研究 AI 与 Agent 技能库

③ 健康管理 / 营养与食品 AI

④ 模型底座与 Agent 基础设施

三、今日趋势观察

四、GitHub 地址速查