2026-11-14 · GitHub AI Daily

WindowBeijing time 2026-11-14 00:00 — 16:30 (GitHub committer time from 2026-11-13 16:00 UTC). Grounded in real commit timestamps (`committer.date`) on each default branch rather than the `pushed_at` approximation, supplemented by same-day releases and official in-repo announcements.

Coverage① AI Coding / programming agents and skill ecosystems ② Medical / clinical-research AI and agent skill libraries ③ Health management / nutrition and food AI ④ Model foundation and agent infrastructure — four lines in total.

De-duplicationCompared one-by-one against every issue published from 2026-10-16 through 2026-11-13 (1905 historical repositories in total); all 155 repositories in this issue appear for the first time. There are no cross-category duplicates inside the four tables, and the index entry count = table row count = unique address count, all three equal. Three repositories that topped last week's trending list — anthropics/skills (★194,243), mattpocock/skills (★323,974) and cloudflare/security-audit-skill (★72,072) — still show no commits on their default branch today, so they are not counted in this issue's tables and are reported honestly in the trends section instead. NVIDIA/SkillEvaluator (★1,154) and addyosmani/agent-skills (★117,271) did make real moves today, but both had already appeared in earlier issues, so under the cross-issue de-duplication rule they are not repeated in the tables; their activity is recorded in the trends section instead.

VerificationStar counts and licenses come from the GitHub REST API (`repos` endpoint and `license` endpoint, with licenses cross-checked against the actual LICENSE file text inside each repository). Commit records come from the `commits` endpoint filtered on `committer.date`; releases come from the `releases` endpoint. Every key README claim was verified by opening the source text, and any repository whose description advertises substantive content (evaluation sets / datasets / benchmarks) had its README read in full to confirm. All repository names, star counts, commit messages and evaluation figures come from actual API responses — nothing is invented.

1. Today's picks (5)

1. Browser vendors start giving "web platform knowledge" an evaluation dashboard: Google Chrome turns 203 guides into an asset that needs continuous regression testing

URL: https://github.com/GoogleChrome/modern-web-guidance-src ｜ ★5154 ｜ Apache-2.0 ｜ A source repository jointly maintained by the Chrome, Edge, Firefox and MDN teams that turns modern web-platform knowledge into a searchable, installable and evaluable skill set for coding agents

What happened

17 commits today, concentrated on two tracks — the evaluation system and the analyzer capabilities: js: add discipline guide, expectations, daily-grind eval, and 11 analyzer rules (#3525), evals: define benchmark lexicon and align ui, cli, and docs (#3562), and analyzing: 4.5x faster legacy-scan (#3565).
The analyzer gained three capabilities, each announced individually: scanning .scss files and blank // comments inside preprocessor style blocks, extracting Lit CSS template literals via the TypeScript AST, and no longer misreading <script-*> / <style-*> custom elements as script or style blocks.
The project's asset strategy is stated plainly: the guides are "designed to be token-efficient", and evals are run "enabling us to prune lowest-common-denominator content that models already know" — content models already know is cut outright.
Delivery is via CLI: npx modern-web-guidance@latest install runs an interactive wizard, and search / retrieve / legacy-scan can each be called on their own. The repository also hosts a public Eval Dashboard.
The README organizes skills by discipline (user experience, CSS layout, accessibility, built-in AI, language detection and so on); it currently lists 203 guide entries.

Why it matters

This picks up exactly where last issue's thread left off — skills are engineering assets that need CI gates. "Prune low-value content models already know, using evals" can be copied straight into my 4P expert-package spec: every intervention-report template I produce today carries a lot of generic nutritional common sense that models certainly already know, while what genuinely deserves context is my own decision framework and threshold table.
The official Eval Dashboard matters a great deal to me: if my 4P platform emits a report without a visible panel showing which skill segments and which rules produced it, I am effectively hiding the most important layer.
Names like "11 analyzer rules", "daily-grind eval" and "benchmark lexicon" show they treat evaluation as a continuous daily pipeline rather than a one-off acceptance test. That is the same discipline as "a failed arm is not an unrun arm", phrased differently.
A sober reminder: this whole stack is maintained jointly by browser vendors and MDN, which means the "general domain knowledge" portion of skills is being absorbed and standardized by the giants. The part of my 4P expert package that cannot be replaced is nutrition-intervention decision logic and clinical boundary rules — not general medical common sense.

2. Clinical knowledge gets an official skill-library shape for the first time: CQF turns CQL, FHIR and clinical reasoning into Agent Skills that ship with evals

URL: https://github.com/cqframework/cqf-skills ｜ ★0 ｜ Apache-2.0 ｜ Open Agent Skills maintained by the Clinical Quality Framework community for authoring, reviewing, testing and packaging clinical knowledge artifacts (CQL libraries, measures, PlanDefinitions and the IGs that publish them)

What happened

21 commits today, every one of them adding a skill and its paired evals: Add usqualitycore and cms-qmd skills, from usqualitycore*, with evals (#10), Add uscore and us-cql-ig skills, split from uscore, with evals (#9), Add cql-ig skill for Using CQL With FHIR 2.0.0, with evals (#8), Add fsh skill (FSH 3.0.0, SUSHI 3.x) with evals (#7), Add fhir-extensions skill for the R4 Extensions Pack, with evals (#6).
The repository currently contains 12 SKILL.md files covering 11 named skills (cql, cql-ig, us-cql-ig, cms-qmd, fhir, fhir-extensions, uscore, usqualitycore, crmi, cpg, fsh), grouped into four installable plugins — cql / fhir / clinical-reasoning / fhir-ig.
There are 114 files in the repository tree related to evaluation, meaning every skill carries its own set of evals rather than one shared suite.
Skill content is pinned to specific specification versions: CQL 1.4 / 1.5.3 / 2.0.0, FHIR R4, US Core 6.1.0 and 9.0.0, US Quality Core 0.5.0, CRMI 2.0.0, and FHIR Extensions Pack 5.4.0. The README explicitly labels the repository status as early development, with skills being brought in one at a time from other repositories.
The skills follow the open Agent Skills format, so they work directly in Claude Code, Claude.ai, the Claude API and any other agent that supports SKILL.md.

Why it matters

This is the sample I have been waiting for: clinical domains now have a community-recognised skill library with version numbers and evals. My homework from the last issue was "the 4P expert package should get a 30–50 item pairwise comparison eval" — CQF supplies three things I can copy directly: ① one eval suite per skill, never a single omnibus eval; ② skills grouped by plugin and individually installable; ③ knowledge version numbers written into the skill names and content (US Core 6.1.0 and 9.0.0 sitting side by side).
It does not solve the question of "who adjudicates clinical correctness". FHIR and CQL have published specs to defer to; clinical nutrition intervention does not — which is exactly why my 4P package must write its own eval set and cannot wait for an upstream solution. Today's evidence supports that judgement.
One honest detail worth recording: the repository labels itself early development. "Official" does not mean "mature" — I should hold my own spec documents to the same standard.

3. An evidence-grounded skill library reaches industrial scale: ASB distils 11,830 metabolomics skills out of methods papers

URL: https://github.com/HolobiomicsLab/asb-skill-collections ｜ ★16 ｜ Apache-2.0 ｜ The public registry of the AgenticScienceBuilder pipeline, distilling published method papers and their code into scientific AI skills and tool records anchored to verbatim evidence

What happened

11 commits today, almost all of them closing gaps in the PII redaction gate, each describing a concrete hole: fix(screen): redact an address whose domain goes on after a hyphen (#162), fix(screen): redact a labelled phone number with a space after the plus (#163), fix(screen): redact an address at any root-zone domain, not only the 35 listed (#161), fix(screen): stop leaving builds out for mentions that name no machine (#160).
The pipeline's core mechanism is spelled out: each skill is written from a published methods paper (journal article or preprint) and, where one exists, its public code repository; it is anchored to verbatim evidence from those sources, annotated with EDAM terms, and passes through licence and PII gates before release.
Scale of the published collections: metabolomics v3.0.2 carries 11,830 skills, 1,989 tools, 20 workflows and 477 papers; alongside it sit genomics (3,246), metagenomics (2,754), epigenomics (2,295), astronomy (1,671), multi-omics (787), metabarcoding (803) and agentic-ai (668) — ten domains in all, each release deposited to Zenodo and pinned to a specific commit and tree digest.
Code licence and content licence are stated separately (Apache-2.0 and CC-BY-4.0); the per-domain repositories are private while the full collection lives in the Zenodo record. The project began at the Dagstuhl Seminar on Computational Metabolomics in April 2026.

Why it matters

This is the one my readers should see most: 11,830 metabolomics skills anchored to 477 papers turns "evidence-based" from a slogan into a checkable pipeline (paper → verbatim evidence → term annotation → licence gate → PII gate → versioned release). If every supplement recommendation in my 4P platform is to carry literature, each one should be pinned in this way: source, verbatim evidence, version.
All 11 of today's commits being PII-gate edge cases shows that the real cost of a gate is not designing it but continuously fixing holes in it. Any "automatically filter sensitive information" flow I build should be expected to need long-term edge-case patching rather than being written once and done.
Separating code and content licences, keeping domain repositories private while shipping full collections via Zenodo — this is a realistic and sustainable way to distribute academic assets, and steadier than cramming everything into one MIT repository.
The most useful point for my one-person company: these skills can serve as a ready-made methodological reference for my GRPM nutrigenomics system and multi-omics intervention work, instead of inventing the "paper to actionable recommendation" chain from scratch.

4. Medical imaging models grow their own MCP surface: MedMLX moves CT / MRI segmentation onto Apple Silicon, locally

URL: https://github.com/MedMLX/medmlx ｜ ★29 ｜ Apache-2.0 ｜ A native medical-imaging inference layer for Apple Silicon, running TotalSegmentator, nnU-Net, MedSAM2, SynthSeg and more through one Python API, CLI and MCP server

What happened

6 commits today, plus two releases — v0.9.1 and medmlx-core v0.2.7 (published 2026-11-13 23:00 UTC).
Today's commits concentrate on tightening interfaces and error classification: Return a typed RunResult from run and wait, Key run outputs by role and keep provider error categories, and Use MedSAM2's pixel-named prompt arguments.
The model catalogue covers real clinical tasks: musclemap (113 muscle and bone structures), totalsegmentator (whole-body CT segmentation), the nnunet family (liver / pancreas / custom PET / BraTS2021), mrsegmentator, siam (whole-head segmentation), synthseg (brain segmentation and parcellation), medsam2 (box / point / mask prompting), sam3 and biomedparse-v2 (text-prompted segmentation).
The three-stage CLI is worth noting: describe lists arguments, readiness checks only Metal and input paths without downloading weights, and run actually executes. medmlx setup --all --progress pre-downloads every model's weights for offline use.
An MCP server and a PyPI package are provided, requiring Python 3.13 and Metal. Model weights download on first use and keep their own separate terms.

Why it matters

This is concrete cost-reduction information for my readers in the health line: running CT whole-body segmentation, organ and lesion segmentation natively on an Apple Silicon machine means imaging-adjacent analysis no longer has to call a cloud API every time. It connects directly to 4P precision nutrition's long-run direction of folding multi-omics and imaging metrics into an intervention plan.
"readiness checks without downloading, weights pre-staged for offline use" is directly useful to me — my local-inference operations experience with Ollama is reproduced here: take an environment and path health check first, then decide whether to pull a large file.
Today's commits on "keying run outputs by role and keeping provider error categories" are textbook engineering discipline: separating errors by source is what lets an upper layer give an explicable failure reason. My 4P platform's OCR path for lab reports should do the same — distinguish "image quality insufficient", "model not installed" and "field parse failed" rather than reporting a blanket "recognition failed".
The Apache-2.0 code licence alongside weights carrying their own terms is worth flagging to readers too: open-source code does not mean commercially usable models, which is exactly the distinction I need to make when attaching a "evidence-based, zero-hallucination" label to content.

5. The DSH ecosystem advances on three fronts at once: Tobs tightens private-Bot authority, DSH gains native ACP, and an enzyme-function-mining microbiome skill package ships as a plugin

URL: https://github.com/BotHarness/DeepSeekBot ｜ ★845 ｜ MIT ｜ Tobs (formerly DeepSeekBot), an open-source GrokBot alternative that runs as a plugin on DeepSeek Harness, where several bots with their own identity, persona and long-term memory work as a team

What happened

95 commits today, one of the highest commit densities in this issue; the latest release is Tobs v1.7.0 (2026-11-12), and the rename from DeepSeekBot to Tobs plus its npm package migration were completed.
Almost every commit today concerns permission boundaries, and all of them move in the same direction — tightening: Preserve Owner bridge composition and close private Bot authority gaps, Merge published Team and private Bot boundaries into E1, fix(client): offer members only a blank Bot to create, Merge pull request #4132: reciprocal Host relationships.
The project runs as a plugin inside DSH. Multiple bots handle research / writing / coding respectively, each keeping its own identity and persona; they can be chatted with privately, pulled into group chats, and @-mentioned for task assignment, with support for Slack, Discord, Feishu / Lark, QQ and WeChat. Bots can create their own scheduled tasks.
Two related DSH repositories also moved today: xintaofei/deepseek-acp (★20, 11 commits) turns DSH into a complete editor-facing coding agent over the Agent Client Protocol; AskTheWay/dsh-auto-memory (★132) turns DSH memory into a four-state trust machine with evidence chains and auditability; Liuyixin-lily/dsh-bio-human-microbiome-function-mining (★0, 10 commits) packages human microbiome function mining as a DSH plugin with 9 built-in skills and 25 scripts, where external tools are executed by the agent only when "the path is accessible and the user allows it".

Why it matters

"Members can only create a blank Bot" and "private Bot authority gaps closed" tell me more than any marketing line does: the real risk in a multi-agent system is not that the model is too weak, it is that permission boundaries are vague. Tobs spent today tightening exactly that line — the mark of this ecosystem moving from toy to usable.
This is the first time in four issues that the DSH ecosystem shows up on three tracks at once — application layer (Tobs, 95 commits), protocol layer (ACP wiring into editors), and methodology layer (a microbiome skill package producing evidence chains and falsifiable experiment plans). Together with dsh-context / dsh-multitask / dsh-as-mcp from 2026-11-13, DSH has moved from "personal plugin" to "industry-scenario adaptation" to "enterprise delivery boundary" — it should be formally promoted to its own tracked line rather than treated as a subclass of agent frameworks.
The microbiome plugin's rule that "whether the agent executes an external tool is decided by whether the path is accessible and the user permits it" is exactly the boundary design my 4P platform should copy: tool execution authority is not decided by the model, but jointly by user authorization and path reachability. That is an order of magnitude stronger than a line of caution in a prompt.
A practical note to myself: I should settle one thing for the 4P platform today — for any step that calls an external bioinformatics tool, the default is "generate the command, do not run it" unless the user explicitly flips the switch. That requires no external capability to adopt.

2. Category roundup

① AI Coding / programming agents and skill ecosystems

Project	Stars	One-liner
addyosmani/agent-skills	117271	Production-grade engineering skills for AI coding agents; 18 commits today including firing debugging on pasted diffs and a plugin manifest bump to 0.6.16
open-gsd/gsd-core	11525	Git-driven "ship it done" core toolchain, 8 commits today
ai-boost/awesome-prompts	9239	Curated prompts, frameworks and papers for LLMs and agents, with an engineering bias
ai-boost/awesome-harness-engineering	5153	Awesome list for AI agent harness engineering: tools, patterns, evals, memory, MCP, permissions
GoogleChrome/modern-web-guidance-src	5154	Web knowledge skill library co-built by Chrome / Edge / Firefox / MDN, 203 guides with a public eval dashboard
ag2ai/ag2	5063	AG2 (formerly AutoGen), the open-source AgentOS; 6 commits today
openBitFun/OpenBitFun	4018	A high-performance agent runtime in Rust with a polished desktop interface; 22 commits today
master-co/css	1957	A markup-driven CSS orchestration language for developers and AI agents
0sec-labs/0	1347	Open-source cybersecurity harness with a web interface for code analysis and penetration testing; 86 commits today
promptdriven/pdd	936	Prompt files are the source, code is the generated output; 94 commits today and a 0.1.32 alpha release
instruktlabs/kiln	327	Build and revise procedural 3D assets with your coding agent via a local MCP server and CLI
yzfly/awesome-context-engineering	228	A curated collection of resources, papers, tools and best practices for context engineering (CC0-1.0)
Ethan-YS/project-brain	174	A folder structure and protocol that let any AI coding session pick up exactly where the last one stopped
wlsdks/ontology-atlas	155	A shared Markdown ontology of the codebase, where missing evidence is marked unknown rather than safe
dengmengmian/CodeLeveler	163	A local-first coding agent that works in your repository and leaves reviewable changes
linuztx/alter-zero	162	A RAM-efficient terminal agent harness for coding, cybersecurity and automation
hwfengcs/DM-Code-Agent	152	An inspectable coding agent: trace changes, diagnose failed runs, share test evidence
AskTheWay/dsh-auto-memory	132	Trustworthy memory for DeepSeek Harness — a four-state trust machine, evidence chains and auditability
nullpointexception-i/agent-sphere	78	Build, orchestrate and run enterprise AI agents in one framework
everruns/everruns	67	An open-source Rust framework and durable runtime for AI agents
Fangyuan025/Chaty	58	A private desktop app that puts the models on your disk to work
dean9703111/auto-edit-workspace	55	Tell Claude Code / Codex / Cursor in plain language to "cut it clean" or "finish the cut" and it does it
maomaochong-ai/open-astravia	53	Open-source local-first AI agent with MCP, skills, plugins, workflows, knowledge base and database built in
nhtera/OxiMux	51	A Rust-native multi-agent development cockpit for macOS — spawn isolated worktrees, run in parallel
rjmurillo/ai-agents	49	Open-source multi-agent toolkit for Claude Code and GitHub Copilot CLI: 30 agents with plan mode
mifunedev/agro	43	AGRO — Agent Governance Runtime Orchestrator, a portable home for autonomous coding agents
tsukhani/jclaw	42	A Java-based enterprise AI harness built in pure Java
hcipengm/cogneva	37	Distributed AI multi-agent autonomous system — digital employees for long-horizon tasks
DotNetAge/mindx	36	A modern AgentOS built for long-running, stateful and multi-agent collaboration
volekerb/hivemill	33	A self-hosted software factory of Claude Code agents with review, QA and a judge merge gate
AmethystLuna/embedded-workbench	32	Embedded C/C++ AI engineering plugin with firmware skills (FreeRTOS, Keil, HardFault, state machines)
Agentic-System-Lab/harness-relay	27	A local group chat where real Claude Code, Codex, Grok, Cursor and Antigravity sessions talk to each other
jcz2020/par	21	PAR — a programmable agent runtime in OCaml 5 with multi-provider LLM support
JoahanMorales/upiixia-agentic-kit	20	Multi-agent coordination kit for Claude Code, Codex, Cursor and Gemini CLI
kongusen/deepstrike	20	An Agent OS microkernel for cross-language agent runtimes
xintaofei/deepseek-acp	20	Turns DeepSeek Harness into a complete editor-facing coding agent over the Agent Client Protocol
fortunto2/solo-factory	18	57 skills, 2 agents and a sensor harness for Claude Code: research, plan, ship and verify
GEMISIS/leviath	14	A structured agent runtime for LLMs: structured context, multi-stage workflows and ECS
Maksim-Burtsev/merl	13	A terminal code navigator: find your way around a project and review the code with an agent
qianshuixia/Shoal	11	A local workbench running Claude Code, Codex and Antigravity CLI side by side on Windows
raultov/knot	11	High-performance Rust MCP server and CLI codebase indexer giving agents deep code retrieval

② Medical / clinical-research AI and agent skill libraries

Project	Stars	One-liner
openmrs/openmrs-core	1994	OpenMRS API and web application code, 5 commits today
ai4protein/VenusFactory2	253	Recursive self-improvement AI agent platform for protein engineering; 20 commits today
MedMLX/medmlx	29	Native medical-imaging inference layer for Apple Silicon — one API / CLI / MCP for segmentation and MedSAM2
ViktorIXAI/claryel-autonomous	14	An AI-built, autonomously evolving harness for local language models with self-testing and validation; 31 commits today
BioSymphony/structure-factory	12	AI-agent toolkit for structural biology: design binders, map structures, screen candidates
lynxkite/lynxkite-2000	12	GPU-accelerated drug discovery, graph analytics and agentic workflows
BioAF/bioAF	10	A self-hosted platform for orchestrating bioinformatics pipelines and managing experiments
aicell-lab/bioengine	9	A distributed AI platform for bioimaging; 22 commits today
donbr/lifesciences-research	9	AI agent wrappers for life-sciences APIs (Open Targets, ChEMBL, UniProt)
WGSExtract/wgsextract-cli	8	A toolkit to extract all sorts of info from your WGS file; 22 commits today
PromptPotter/prompt-potter-optimizer	7	LLM-driven evolution of prompts and pipelines, evolving both the prompt and its parameters
BioSymphony/small-molecules	6	Agent skill for choosing small-molecule design tools for molecular generation and synthesis
MEFORORG/MessageFoundry	5	Healthcare interface engine; 26 commits today
Wulfgardr/mediflow	4	Free open-source management of clinical records, documents and outpatient activity; 26 commits today
HarperZ9/flywheel	3	Run AI tasks on any hosted or local model, then check answers against the sources that decide
fhirlint/fhirlint	3	FHIR validator CLI with readable output, watch mode and CI reports; 22 commits today
jibeomko/PanIsoGuard	2	Caller-agnostic adjudication of long-read RNA-seq novel isoforms
1999labs/RTR-bench	1	A benchmark of whether language models can read the room; 17 commits today
fishhrbio/Proteonator	1	Local protein structure and precursor workspace for Windows, macOS and Ubuntu
naiyixi/PureScience	1	A research workbench that keeps every step of your science inspectable
poincare-lijiashu/medical-agent	1	Hospital-scale clinical collaboration platform 2.0: 16 domain services, 913 HTTP endpoints, FHIR R4
LaBiOmicS/q2-periobiome	0	QIIME 2 plugin for Socransky complexes and dysbiosis indices on subgingival 16S rRNA data
MHKhan50/agentic-ai-research-log	0	Daily reproducible experiments on LLM agents, clinical NLP and local models (Ollama)
Maggydata/sovereign-health-assistant	0	Fully local AI assistant for sensitive medical documents: open-source LLM (Ollama) plus PII handling
NgHuuGiau/OncoVision	0	Cancer analysis system; 12 commits today
Raouka/healthpoint	0	Local-first clinical interoperability and reconciliation; 16 commits today
Rebell-Leader/BlisterKompass	0	A printable picture sheet telling older people which pill is which; 13 commits today
ananthv-94/medical-document-ai-agent	0	Multimodal medical document intelligence agent using LangGraph, RAG, OCR and LLMs
cognifityai/verdict	0	Open-source observability, evaluation and drift monitoring for AI agents and LLM applications
cqframework/cqf-skills	0	The official skill library for CQL / FHIR / clinical reasoning — 12 skills, each with evals
dzulfikar08/irit	0	Lean bioinformatics tools for ordinary laptops: sample sheets, read QC and reference datasets
james-merithew/openemr-red-team	0	Four-agent red-team platform (orchestrator, attacker, judge, documentation) running continuous adversarial tests
jonperron/fhir-linter	0	A FHIR validator and linter written in Rust; 32 commits today
mgegen15-dev/For_John	0	Open research on osteosarcoma: overlooked leads, lost trial results and untested drug targets
r-morenov/clinical-guideline-rag	0	Local RAG plus agent platform over public clinical guidelines (CKD, diabetes, hypertension)
studiome/altocumulus	0	Clinical ledger for patients, diagnoses, surgeries and hospitalizations, with elective-surgery thresholds
supermax01/fireboat	0	Gives AI agents safe, read-only access to one patient's FHIR health records over MCP
thepiglab/fhir-agent-demo	0	FHIR agent demo environment; 17 commits today
thylinao1/bayes-cot-faithfulness	0	Bayesian causal-mediation framework for measuring whether a model's chain-of-thought is faithful; 22 commits today
openmrs/openmrs-esm-patient-chart	96	OpenMRS 3 patient chart frontend module family; 6 commits today

③ Health management / nutrition and food AI

Project	Stars	One-liner
YunYouJun/cook	6547	Chinese home cooking — recipes plus a cooking assistant; 13 commits today
nextcloud/cookbook	654	A self-hosted library for all your recipes (AGPL-3.0)
marcionicolauvip1/cookdiary-desktop-planner	102	Windows recipe planner with weekly menus and step-by-step guidance
reaper47/recipya-rs	41	A clean, simple and powerful recipe manager (AGPL-3.0)
googlarz/suunto-mcp	40	MCP server connecting Suunto watch data to Claude and other AI assistants; 23 commits today
megabyte0x/healthysync	17	Privacy-first iOS Apple Health sync for private APIs and self-hosted backends
HolobiomicsLab/asb-skill-collections	16	Evidence-grounded scientific skill library distilled from methods papers — 11,830 metabolomics skills
suyiiyii/alldayrec	16	All-day recording so agents understand you better: keep the original audio, speaker ID, summaries, todos and MCP
sergiparpal/meal-manager	10	Hermes agent meal-planning plugin: a deterministic core under a conversational shell, where the LLM only translates
marco308/meals	5	Meal options planner with an aisle-sorted shopping list; 29 commits today
RDA-MOMSI/Dashboard	4	MOMSI Standard Landscape Review curation workflow and multi-omics dashboard
plyush1993/MetaboCensoR	3	Shiny app for filtering redundant features in LC-MS
drew2323/pap-analytics-mcp	2	Device-agnostic PAP (CPAP/APAP/ASV) therapy analytics for AI agents via MCP
kirmalyshev/calodex	2	eait.fit — a photo-first nutrition coach: backend, client contract, web app and Telegram bot
savorum/savorum	2	A recipe library you can keep: save, organize, plan, cook and print; 23 commits today
BatterWorks/Kamosu	1	A self-hosted cookbook where every version is preserved
BenCBanks/kinetta	1	Workouts and food in one app: activity tracking, training plans and food logging
api-evangelist/foodvisor	1	An independent third-party profile of a public API surface
h0tman5/leap	1	Whoop-style recovery, sleep and strain tracking for the Amazfit ecosystem
jtremblay/microbiomeutils	1	Python utility to generate distance matrices, run PCoAs and produce taxonomic summaries
trOOnies/pamela	1	Personal physical-activity and macro tracker app
Bao992/NutriTrack	0	An autonomous AI-powered school nutrition operations colleague
EssaHusary/CSC780_TermProject	0	Nutrition label interpreter — photograph a label and it parses it
Liuyixin-lily/dsh-bio-human-microbiome-function-mining	0	Human microbiome function-mining dsh plugin: 9 skills and 25 scripts producing evidence chains and experiment designs
Nahlah-CS/Gut-microbiome-colorectal-analysis	0	16S rRNA analysis of gut bacteria in healthy, polyp and colorectal-cancer samples
OpenVibers/OpenVibe.Food	0	Food banks, budget grocers, meal plans and cupboard recipes for the household
Significant-Hobbies/food	0	Food reference for Indian shoppers: compare nutrition labels and trace source evidence
SprockettLab/MAGmaker	0	Assembling metagenome-assembled genomes from metagenomic data
colincooperio/garmin-mcp	0	Ask Claude, ChatGPT or any AI assistant about your Garmin data via Garmin's official API; 9 commits today
deimosowen/IronPulse	0	Strength-training timer for Huawei Watch 4 Pro with live heart-rate summaries
interias/BISS	0	Self-hosted calorie and macro tracker: scan a barcode, snap the nutrition label, log meals
johnfoster1012-pixel/supplement-intelligence	0	Evidence-based research notes for supplements — ingredient analysis and evidence curation
zoelsner/qook	0	iOS meal-planning app with AI-generated recipes, weekly planning and a grocery flow
ReceiptCal/open-nutrition-transaction	1	OpenNutri, an open standard letting nutrition data travel with the food you buy, with provenance and consent boundaries
ReceiptCal/receiptcal-plugin	1	Open-source MCP plugin bringing ReceiptCal food data into agents
cjm-m/wechat-pantry-ai	1	A WeChat mini program for a shared fridge plus AI meal planning, with a household ReAct agent

④ Model foundation and agent infrastructure

Project	Stars	One-liner
infiniflow/infinity	4765	The AI-native database for LLM applications, with very fast hybrid search
DataDog/datadog-agent	3771	Main repository for the Datadog Agent; 33 commits today
truera/trulens	3639	Evaluation and tracking for LLM experiments and AI agents
doobidoo/mcp-memory-service	2126	Open-source persistent memory for AI agent pipelines (LangGraph, CrewAI, AutoGen) and clients
ArcadeData/arcadedb	1251	One multi-model DBMS covering SQL, Cypher, Gremlin and HTTP/JSON
basetenlabs/truss	1235	The simplest way to serve AI/ML models in production
BotHarness/DeepSeekBot	845	Tobs on DeepSeek Harness: a multi-bot team with tightened private-Bot authority
jeffhajewski/latticedb	755	Embedded single-file knowledge-graph database with vector search and full-text search
matrixorigin/memoria	627	Secure memory management for AI agents, ensuring data integrity and reducing hallucinations
labring/aiproxy	570	High-performance AI gateway speaking OpenAI, Claude and Gemini protocols
ratel-ai/ratel	561	The self-improvement layer for AI agents, reading traces to find what failed
ome-projects/ome	525	Open Model Engine — a Kubernetes operator for LLM serving, GPU scheduling and model operations
vortico/flama	307	The production framework for predictive and generative AI, serving any model as an API
relatedb/relate	257	An embedded local agent-native graph database supporting Cypher 25, written in Rust
arconia-io/arconia	199	A framework that acts as an add-on for Spring Boot to enhance model context
YoKONCy/TriviumDB	191	A lightweight embedded database engine fusing vector search, graph and documents natively
tma1-ai/tma1	124	Local-first observability your agent reads back: every LLM call recorded and routed
wso2/agent-manager	117	An open control plane for enterprises to deploy, manage and govern agents
techysy/10router	108	Local AI routing gateway and dashboard
Siddhant-K-code/agent-trace	106	Local evidence and replay for coding agents: see what happened and what the trace contains
MarcoPorcellato/matryca-plumber	102	Local-first AI daemon for Logseq OG: background semantic indexing, link hygiene and age checks
B3hnamR/Polygate	27	A free model gateway for Windows and Linux with WARP, proxy pool and Relay
OpenRouterTeam/benchmark-harness	26	OpenRouter's TypeScript harness for reproducible LLM benchmarks and evaluations
jersobh/epochdb	24	An agentic memory engine built for lossless, tiered verbatim storage and multi-hop retrieval
mirror-mind-ai/mirror	21	A local-first memory and identity framework for agentic AI runtimes
ROCm/rocm-aic	19	AMD Infinity Context — a distributed, shared, low-latency KV cache solution
visorcraft/MongrelDB-V	11	Pure V client for MongrelDB with SQL, vector search and full-text search
LatentPort/latentport	10	Programmable LLM state: transfer, compact, save and fork KV caches and recurrent memory
tapiruslab/TapirusDB	9	A 100% safe-Rust embedded quad-model AI database and cognitive memory engine
visorcraft/MongrelDB	9	A fast embedded-plus-server columnar database with SQL, vector search and full-text search
BeppeTemp/cartographer	67	A shared, git-backed wiki your team's AI agents maintain
Dami-wq-k03/nano-lmcache	28	An educational, step-by-step implementation of an LLM KV cache
ai-decentralized/Difflet	10	Fast diffusion-transformer inference across accelerators
agent-guide/agent-gateway	19	A self-hosted AI gateway for governed LLM, MCP, ACP and A2A traffic
ClickHouse/loom	23	Shared long-term memory, document search and a code index; 40 commits today
logicplanes/kaeso	20	A self-hosted AI gateway in Rust speaking OpenAI, Anthropic and other protocols
sattyamjjain/agent-airlock	18	A type-checker and contract layer for AI agent tool calls
ReliablyObserve/loki-vl-proxy	50	A Loki-compatible API proxy for VictoriaLogs, usable from Grafana

3. Trends today

⭐⭐⭐ "Skills" today showed up in three institutionalized shapes at once: vendor-maintained, community-maintained and paper-pipeline-derived. Chrome/Edge/Firefox/MDN maintain 203 web guides with an evaluation dashboard; the CQF community maintains 12 skills each with its own evals; ASB distils 11,830 skills anchored to 477 methods papers. What the three share is not content but mechanism: a skill must be independently evaluable, must carry a version number, and must be traceable to a source. Those are exactly the three things my 4P expert package still lacks — the sharpest lesson of this issue for me.
⭐⭐⭐ "Missing evidence" is now modelled as a first-class state rather than silently ignored. All 11 of ASB's commits today close concrete holes in the PII gate (email domains after a hyphen, a space after a plus sign, any root-zone domain beyond the 35 listed); ontology-atlas states explicitly that "missing evidence shows as unknown, never as safe", and rules that Markdown examples, hidden comments and blockquotes cannot support capability claims; every cqframework skill names the spec version it relies on (US Core 6.1.0 and 9.0.0 side by side). Three unrelated projects converging on the same principle: never write "not checked" as "no problem". This belongs directly in my spec-document template.
⭐⭐⭐ Permission boundaries are tightening across the multi-agent ecosystem today. Tobs's commits all move in the tightening direction: closing private-Bot authority gaps, letting members create only a blank Bot, preserving the Owner bridge composition. The real risk in a multi-agent system is not that the model is too weak, it is that permission boundaries are vague — the same rule clinmesh implemented last issue ("the final Actor is always the logged-in human clicking the review box"), arriving at a different layer.
⭐⭐ "AI picks, code computes" picked up two more implementations worth copying. meal-manager confines the LLM to being a "semantic translator" and states plainly that "The LLM is a Translator, Not a Judge" — scoring, inventory and session state are all deterministic Python, and ambiguity is resolved exactly once at the parsing edge; kawsay (which entered yesterday's issue) likewise lets the AI pick only catalogue items and portions, with every number computed from a table. Seeing the same principle converge independently across two ecosystems in consecutive issues suggests this is not one team's preference but the optimal shape for this class of product.
⭐⭐ Local inference is expanding from text models to specialist models. MedMLX moves CT/MRI segmentation, MedSAM2 and SynthSeg onto Apple Silicon and exposes an MCP surface with offline weight staging; last issue's ddalcu/mlx-serve pointed the same way. For this site the meaning is concrete: if 4P is to ingest imaging metrics, running them locally is moving from "unrealistic" to "a path that exists".
⭐⭐ Last issue's watchlist produced results today, but none of them enter the tables. NVIDIA/SkillEvaluator made 8 commits, including feat(harness): flag a harness release the profiles were not verified on (#545) and fix(scanners): enforce pinned SkillSpector during automatic selection (#550), continuing the "explicitly mark unverified combinations" thread; addyosmani/agent-skills made 18 commits (plugin manifests bumped to 0.6.16); but anthropics/skills, mattpocock/skills and cloudflare/security-audit-skill still show no commits on their default branches today. All had appeared in earlier issues, so under the cross-issue de-duplication rule they are not repeated in the tables.
⭐ The DSH ecosystem ran three tracks in parallel today — the most worth tracking on its own line. Application layer (Tobs, 95 commits), protocol layer (deepseek-acp wiring DSH into editors via ACP) and methodology layer (dsh-bio-human-microbiome-function-mining producing evidence chains and falsifiable experiment designs from 9 skills and 25 scripts). Together with dsh-context / dsh-multitask / dsh-as-mcp on 2026-11-13, DSH has gone from "personal plugin" to "industry-scenario adaptation" to "enterprise delivery boundary" in one week, and is now formally promoted to an independently tracked line.
⭐ A practical note for the one-person company: of 155 entries, the ones actionable next week still number about three (CQF's one-skill-one-evals organization, OpenNutri's practice of tagging every value with its source and confidence, meal-manager's "LLM only translates" boundary). Everything else is long-run observation — do not let yourself be dragged into "reading 155 repositories a week"; landing one per week is enough.

4. Repo index

GitHub AI Daily · 2026-11-14 (Saturday)

1. Today's picks (5)

1. Browser vendors start giving "web platform knowledge" an evaluation dashboard: Google Chrome turns 203 guides into an asset that needs continuous regression testing

2. Clinical knowledge gets an official skill-library shape for the first time: CQF turns CQL, FHIR and clinical reasoning into Agent Skills that ship with evals

3. An evidence-grounded skill library reaches industrial scale: ASB distils 11,830 metabolomics skills out of methods papers

4. Medical imaging models grow their own MCP surface: MedMLX moves CT / MRI segmentation onto Apple Silicon, locally

5. The DSH ecosystem advances on three fronts at once: Tobs tightens private-Bot authority, DSH gains native ACP, and an enzyme-function-mining microbiome skill package ships as a plugin

2. Category roundup

① AI Coding / programming agents and skill ecosystems

② Medical / clinical-research AI and agent skill libraries

③ Health management / nutrition and food AI

④ Model foundation and agent infrastructure

3. Trends today

4. Repo index