GitHub AI Daily · 2026-11-03 (Tuesday)

1. Today's picks (5)

1. ClawBio/ClawBio: writing a domain expert's analytical decisions into a versionable, reviewable, reproducible machine contract — the way this skill library is organised is precisely the layer my GRPM nutrigenomics system is missing

URL: https://github.com/ClawBio/ClawBio ｜ ★1,172 ｜ MIT (per the README; the API reports NOASSERTION, see Verification) ｜ Python ｜ a bioinformatics-native AI agent skill library — local-first, reproducible, with a CLI / Python API / Claude Code plugin / OpenClaw Gateway / Telegram & Discord entry points

What happened

Three commits inside the window (11-02 08:49Z ~ 18:41Z). The volume is small, but every one lands on governance rather than features.

• The substantive work this issue is automated site-stat updates plus a fix to a skill document's length: chore: update site stats (automated) (contributor, test-count and fork-count figures maintained automatically), plus the merge of fix/issue-731-igv-validator-skill-md-over-5-000-words — a SKILL.md over 5,000 words is judged non-compliant, and that limit is now enforced in CI.
• Skill maturity is quantified into six levels and written into a machine-readable skills/catalog.json: spec-only (only a SKILL.md) → scripted (runnable code) → tested (has tests) → cli-registered (callable via clawbio run) → ci-validated (explicitly run in CI) → bench-validated (backed by a blocking scientific benchmark). "Which level has this skill reached" is a queryable fact, not a self-description in a README.
• Every skill output ships a reproducibility/ directory: commands.sh (the replay command for this run), environment.yml (a Conda environment snapshot), checksums.sha256 (SHA-256 of selected output files) and runtime-lock.json. In other words: any analytical conclusion can be recomputed offline by a third party.
• Licensing is split into three fields: each skill declares license / model_license / data_license separately, and the strictest of the three governs. It also states explicitly that "empty means 'none declared', never 'permissive'" — which is precisely where selection goes wrong most often.
• The data-egress list is enforced in code: docs/data-handling.md lists by five categories which skills send data outward (downloading reference data only / sending only query terms / sending variants / sending full sequences / sending chat content), recording for each the target host, trigger condition, exactly what is sent, the credential variable and the offline behaviour; the tests scan skill code for outbound calls, and a networked skill with no registry entry fails CI outright.

Why it matters

• This repository is almost exactly the target shape of my GRPM nutrigenomics system. What I am building today: extract 40 nutrition-relevant SNPs from a 2,526-page gene report and generate personalised supplement recommendations. ClawBio's PharmGx Reporter is the general-purpose version of the same task — it accepts consumer genetic data from 23andMe and AncestryDNA (with automatic gzip detection), extracts 31 pharmacogenomic SNPs across 12 genes (CYP2C19 / CYP2D6 / CYP2C9 / VKORC1 / SLCO1B1 / DPYD / TPMT / UGT1A1 / CYP3A5 / CYP2B6 / NUDT15 / CYP1A2), identifies star alleles, determines metaboliser phenotypes, and looks up CPIC guidance for 51 medications — zero dependencies, under 1 second. It separates "the expert's analytical decision" from "the executing code" into independently versioned artefacts, which is exactly the specification layer I need to add to GRPM's productisation (upload → analyse → export → share).
• Its negative lessons are worth more than its features. ClawBio publicly corrected one of its own scorecards: the site had published 168/182 (92.3%), while a continuous run actually produced 161/182 (88.5%) — because the leaderboard workflow failed when GitHub Actions could not open a pull request, and nobody noticed the statistics page had been stale for 87 days; one regression source was nutrigx-advisor dropping from 10/10 to 0/10. "An automatically updated statistics page can quietly stop working" — I am noting this because my own intervention-report page also carries auto-generated statistics.
• It deleted its own "first" claim and added a CI guard against regression. Integrity and governance sweep (M1) explicitly says "remove the 'first bioinformatics-native AI agent skill library' primacy claim", removing it from the README, pyproject.toml, the marketplace page, the variant-interpretation workshop page and the Bioconda recipe, and adding a CI gate so the sentence cannot reappear. A project willing to put its own promotional language under regression testing has credibility that can be trusted.
• It retired a feature for the right reason. soul2dna (turning a fictional character profile into a genome and "breeding" offspring by Mendelian inheritance) was pulled from the skill catalogue because "there is no scientific basis for mapping traits of character to genotypes", taking the catalogue from 97 skills to 96. A 1,172-star project killing its own coolest feature for the sole reason that "this does not hold up scientifically" — that standard of judgement is worth more than any technical highlight.
• On selection: MIT (per the README); at the skill layer there is one GPL-3.0 and one PROPRIETARY, but the bulk is commercially usable. Note also that it no longer offers an MCP server — since 0.7.0 it was deprecated and in 0.8.0 it was removed. Access is now via CLI / Python API / Claude Code plugin / Agent Skills directory (copy into ~/.agents/skills/).

2. cortexkit/aft: giving an agent a real IDE and OS — swapping "read the file, change the code" for "read and write by symbol", and keeping a restorable backup behind every change

URL: https://github.com/cortexkit/aft ｜ ★334 ｜ MIT ｜ Rust ｜ the "sensorimotor cortex" for coding agents (OpenCode / Pi / oh-my-pi), built by hoisting the host's built-in tool slots plus an aft_ family of tools

What happened

253 commits inside the window — the highest commit count of any repository this issue, and the only infrastructure project to make the picks.

• The architecture is framed as a "sensorimotor cortex" with three layers. The sensory cortex (perceive) provides aft_outline (every symbol in a file, directory or remote URL, with its kind, name, line range, visibility and nested members), aft_zoom (drill into one function, class or type, optionally with callgraph: true), aft_search (semantic + lexical hybrid retrieval over an indexed codebase, with local, OpenAI-compatible or Ollama embedding backends), aft_callgraph (callers, callees, data flow, impact analysis and the shortest call path between two symbols across the workspace) and aft_inspect (a one-call codebase health report covering LSP errors and warnings, TODOs, metrics, dead code, unused exports and duplicates).
• The motor cortex (act) provides edit (fuzzy find/replace, or replacement of a named symbol directly), write (auto-created directories, backup, formatting and optional inline diagnostics), apply_patch (multi-file *** Begin Patch with atomic rollback), aft_import (language-aware import add, remove and organise) and ast_grep_search / ast_grep_replace (structural search and replace with AST patterns and meta-variables).
• The brainstem (the OS layer) provides bash (command rewriting cat→read, per-command output compression, tree-sitter permission scanning), background tasks (bash_status / bash_kill / bash_watch, with tasks and their completion state surviving restarts), multi-tier output compression (turning firehose CLI output from test runners, installers, docker ps and kubectl into the few lines that matter while keeping errors and summaries), real interactive PTY sessions (able to drive a REPL or even a nested agent), and aft_safety's per-file undo stack, named checkpoints and restore — every edit is backed up to disk and survives both bridge and host restarts.
• Today's commit thread is "give every edge case an explicit answer": TODO scan bounds: mark the skipped-file reason as approved (a skipped file must state its reason and be marked as approved), callgraph read bound (a bound on callgraph reads), mason: preserve Windows task metadata durability on equal writes (Windows task metadata must still be durable on an equal write), mason: classify planned graph loads by typed state (planned graph loads distinguished by typed state) and mason: preserve fixture attribute policy in acquisition workers.
• When a list is truncated it must say so: replies end with shown N of M <unit> (<reason>) · narrow: <knobs> — the absence of that line means the list is complete.

Why it matters

• "No such line means the list is complete" is a rule I can move across immediately. Every time my 4P precision-nutrition platform asks an LLM for an intervention report, if the 24 risk dimensions come back truncated without saying so, I will assume "that's all of them". Making an agent self-certify the completeness of its output costs almost nothing and eliminates an entire class of silent errors.
• The undo stack plus named checkpoints is exactly what I lack when tuning prompts. In the previous issue (11-02) I recorded agent-stash's "save the change as a switchable version" — today this project makes the same idea into an infrastructure layer: aft_safety provides a per-file undo stack and named checkpoints, backing every change up before it hits disk, with the backups surviving bridge and host restarts. When I tune the prompts behind a 4P supplement plan I routinely overwrite the original and afterwards cannot say which version produced a given conclusion — this layer closes that gap.
• Semantic search plus a call graph is the right answer to "the codebase is too big and the agent only ever reads whole files." aft_search finds code by meaning, aft_callgraph tracks impact. "Impact analysis" applies just as well to the supplement platform — change one algorithm and I need to know which intervention-report templates move with it.
• Its output-compression strategy is worth copying outright. "Keep the errors and the summary, drop the noise" is exactly what I want when slimming the GLP-1 nutrition companion's reports: what the user needs is "which markers are out of range and what to adjust", not all 24 dimensions listed.
• MIT licensed — readable, modifiable, commercially usable. For a one-person company this is a candidate to evaluate as a foundation. Note it is a Rust binary with thin adapters (OpenCode / Pi / oh-my-pi) and does not target Claude Code directly.

3. zilet/cairn: reading labs, training, food and sleep as one picture, then offering only the one thing worth doing today — and "no scores, no nudges" is its most counter-cultural sentence

URL: https://github.com/zilet/cairn ｜ ★11 ｜ MIT ｜ TypeScript ｜ a self-hosted wellness OS — PWA + REST + MCP, with all data in a single SQLite file you own

What happened

Six commits inside the window (11-02 07:15Z ~ 13:47Z), with the main thread being "from reading you, to deciding for you".

• This issue pushes "finished goals are looked back on, and a race's lessons shape the next build": feat(horizon): finished goals are looked back on, and a race's lessons shape the next build; and feat(horizon): a finished race is said back — result, recovery days, and what's next — a completed race is reported back with its result, recovery days and what comes next.
• Two bugs about "misleading readings" were fixed, and both matter: fix(day-read): a session still being logged is not the day's work done (a session still being logged does not count as the day's work being finished); and fix(week): a race effort reads as itself, and a health hold's wording agrees in number (a race effort must be read as it actually was, and a health hold's wording must agree with its number).
• Setting a calorie target moved into the conversation: feat(chat): set the calorie target from chat, released as v2.1.11.
• A flagged lab marker propagates by itself into meals, training and the watch, each carrying a "why" and a citation; records are grouped in clinical order with the lab's own flag kept apart from the optimal band; and the doctor packet builds itself — findings, your questions, body composition, results grouped by panel, with toggleable sections, a preview and something you can hand over.
• Expenditure comes from your real weight trend: "A thin logging week lowers confidence instead of scolding" — a thin week of logging lowers confidence rather than scolding you.
• The product boundaries are written into the README: "no scores and no streaks. Nothing is pushed at you. Every read is a suggestion; you drive. Anything clinical still waits for you to say yes."

Why it matters

• "No scores, no nudges" lands exactly on my 4P platform's current problem. In the previous issue I recorded Calibrate's 2×2 dependence matrix and concluded that "presenting 24 risk dimensions at equal weight is the same as showing no risk prompt at all." Cairn goes further: no scores, no streaks, nothing pushed at you — each day it offers one thing worth doing, while a team behind it (strength, endurance, nutrition, physio, and an informational-only health seat) each names the next step toward your milestones; bounded, reversible changes land on their own, with the why and a one-tap way to put them back.
• "A bounded, reversible change lands with its reason plus a one-tap undo" is the interaction paradigm AI assistants should follow when they enter health decisions. My supplement platform today either leaves every step to a human to execute (slow) or lets the AI change the plan directly (dangerous and untraceable). Cairn offers a third path: bounded, reversible changes take effect automatically, with a reason and an undo; anything clinical waits for the user to say yes. This could map almost verbatim onto the "suggestion execution layer" of my 4P intervention report.
• "A thin week of logging lowers confidence instead of scolding the user" is also a compliance point. My platform today is "either there is data or there is no recommendation" — the middle state is missing. Confidence grading is the key mechanism that lets a system operate safely when data is thin, and it maps naturally onto what I recorded last issue as "insufficient evidence must be an independent conclusion rather than folded into a pass".
• All data sits in a single SQLite file you own — zero cloud, zero accounts. What I run is a local-first 4P precision-nutrition platform (Flask + SQLite), so the architectural choice is identical; Cairn additionally exposes REST and MCP, which is a ready reference for the 4pwellness.org.cn MCP connector line.
• It explicitly supports a "coaching agent" using Claude Code / Codex / Antigravity / Grok, installed into its own persistent home volume, logged in with your own account, with no shared key. That matches both my local dsh ecosystem and the WorkBuddy model — the agent is pluggable and the key always stays with the user.

4. grotyx/Academic_writing_c_claudecode: when the evidence only supports "association", the wording is not allowed to escalate to "causation" — that rule belongs straight in my content guidelines

URL: https://github.com/grotyx/Academic_writing_c_claudecode ｜ ★35 ｜ unstated (the core endpoint confirms there is no LICENSE file) ｜ Python ｜ ManuWright — an evidence-bound medical manuscript engine for AI agents (Claude Code / Codex / Gemini)

What happened

Seven commits inside the window (11-02 20:50 +09:00 ~ 11-03 11:22 +09:00) — three releases in a single day.

• v1.10.0 (11-02 22:50): adds manuwright bind (result-binding proposals), phrase-level numeric exemptions, tracked-changes revision DOCX and journal P-value notation.
• v1.10.1 (11-03 00:51): manuwright figure — figures are drawn straight from the results CSV and carry provenance.
• v1.10.2 (11-03 11:22): figure style presets (prism default, nejm, grey) and denser log-axis ticks.
• The core mechanism is a two-stage "plan first, write second" gate: the drafts/draft_plan.md plan file must be completed first, and the plan check counts ticked options and [EVID:] lines, accepts "None" for an evidence gap, and names missing sections in words.
• Citation verification is hard-edged: the scripts/search_pubmed.py audit covers weighted metadata matching, DOI mismatch detection, and retraction / concern / erratum checking; author names match on whole words and do not require accent characters to agree, years compare numerically, and PubMed retrieval runs in batches of 200.
• Claim strength is graded on verbs, not nouns or adjectives: check_claim_strength.py rates cited sentences, and honours the Claim Strength and Allowed Wording that each entry in evidence.md carries. The key rule is "explicit association caps a sentence at associative" — an explicit association caps a sentence's conclusion strength at associative.
• Blind review and revision consistency: when re-reviewing a revision, the response letter is withheld first so the reviewer forms an independent judgement; if the reviewer raises their verdict after reading the reply, a matching final_anchor must exist in the revised text. Review comments are refused if unnumbered.
• Style comes from measured journals rather than from a prompt: the reference corpus is 33 openly licensed original-research papers (JAMA Surgery, JAMA Network Open, Lancet, BMJ, Nature; 2019–2022), and the repo stores only aggregate numbers, generic multi-journal phrases, source DOIs and licences — no paper text. Personal style learning requires at least 3 papers and 30 sentences, and never falls below the journal p90.

Why it matters

• "Explicit association caps a sentence at associative" is the guardrail my WeChat public-account content needs. I produce evidence-bound content (the "4P Precision Nutrition Watch" and "Health 100 Observation" WeChat public accounts) and have always required "zero AI hallucination, references mandatory". But "has a reference" and "the strength of the conclusion does not exceed the strength of the evidence" are two different things — an article citing a cross-sectional study and claiming an intervention "improves atherosclerosis" has a real reference and an out-of-bounds conclusion. ManuWright makes this an executable check: each piece of evidence in the ledger declares the strongest wording it can support, and over-claiming raises an error.
• "Accepting 'None' as a declared evidence gap" maps directly onto my GRPM reports. In my nutrigenomic recommendations, when the evidence for a given SNP is thin, my current behaviour is "give no recommendation". Changing that to "state explicitly that this target's evidence is insufficient" is more useful — what matters is making "missing evidence" a declared fact rather than a silent blank. This is exactly consistent with what I recorded last issue as "insufficient evidence must be an independent conclusion".
• "Figures drawn directly from the results CSV with provenance" (v1.10.1) is a ready-made answer for my intervention-report charts. My GLP-1 nutrition companion has lab-marker trend tracking, and the charts are currently drawn by the front end — generating figures from the persisted result data with provenance eliminates the whole class of "the number in the chart doesn't match the database" problems.
• Retraction / concern / erratum checking is the most commonly missed link in evidence-bound content. When I cite literature I normally verify only that the paper exists, not whether it has been retracted. ManuWright folds it into the routine audit flow.
• The triple threshold for style learning (3 papers + 30 sentences + never below p90) is worth copying. One long-standing frustration of mine: I want AI to keep my personal style without learning a template voice. "At least 3 papers and 30 sentences before it may be enabled, and never below the journal p90" gives 'learn my personal style' an executable admission threshold and a ceiling.
• One licensing fact that must not be missed: this repository has no LICENSE file. So its methodology can be read and its thinking borrowed, but the code cannot go directly into a product I intend to sell. Commercial use would require contacting the author first.

5. pmanko/clinical-ai-validation-harness: a single aggregate score is not clinical validation — it preserves the whole evidence trail behind every judgement so a reviewer can inspect it themselves

URL: https://github.com/pmanko/clinical-ai-validation-harness ｜ ★3 ｜ unstated (the core endpoint confirms there is no LICENSE file) ｜ Python ｜ a clinical AI validation harness: run experiments, keep evidence, produce reviewable reports (Python 3.11+ / uv)

What happened

Five commits inside the window (11-01 21:34Z ~ 11-02 05:38Z), all circling the question of how to make people trust a result.

• The project's opening sentence states the position outright: "The goal is evidence a reviewer can inspect: which records supported an answer, which SQL actually ran, what failed, and why a reviewer accepted or rejected a result. Aggregate scores alone are not clinical validation."
• The substantive work this issue introduces a three-way label contract and "judge cards": Score every check against the judge and read the three-way label contract (every check is scored against the judge and the three-way label contract is read), Judge cards, repeated passes and their agreement (judge cards, repeated passes and the agreement between them), WAHS 4-Use-Case Evaluation Runner, Hard-Subset Stratification & Condenser Mode Matrix (W4, W12) (a four-use-case evaluation runner, hard-subset stratification and a condenser mode matrix), and Drop the live QueryStore reader from corpus alignment (the live QueryStore reader is removed from corpus alignment).
• Even the display rules are treated as a specification: Show each answer and In-Depth the way the product does, with the checks that marked it — the report must present answers and "in-depth" content the way the product actually does, with the checks that flagged them attached.
• Every run drops a complete evidence packet: run_manifest.json, events.jsonl, plus results.jsonl, captured inputs, traces and review records; provenance is either supplied by the caller or observed from the target — "unknown revisions/model details are recorded as missing, not substituted with workspace pins".
• Offline regeneration is an explicit requirement: a report must use the inputs captured at the time, not current product sources, live services or the author's original dataset directory; the complete run directory must be copied as a whole.
• The bar for changes is written into CONTRIBUTING: behavioural changes need focused tests, diverse failure and abstention cases, versioned metadata and record-level evidence; "accepted data mappings/transforms are reviewed and deterministic; LLM proposals remain advisory"; and any material change to model, prompt, retrieval, mapping or pipeline requires a change record describing the validation protocol, impact and residual risk.

Why it matters

• "Aggregate scores alone are not clinical validation" points straight at my 4P platform's soft spot. My intervention report has a 9-chapter structure, 24 risk dimensions and 6 risk levels — which is essentially a total score plus its breakdown. What this project offers is not another scoring scheme but the list of "what else must be kept besides the score": which records supported the answer, which SQL actually ran, what failed, why the reviewer accepted or rejected it. Moving this set across to the 4P platform is the complete version of the "intervention-report persistence checklist" I recorded last issue.
• "Unknown revisions/model details are recorded as missing, not substituted with workspace pins" is a life-saving rule. I hit this constantly when producing reports: the model changed, the prompt changed, but the conclusion in the report stands, and there is no way to tell which version produced it. Preferring "unknown" over filling a historical gap with today's value is the mirror image of the yantrikdb footgun I recorded last issue ("a second SQLite instance in the same process silently corrupts storage"): never fill an uncertain hole with something that merely looks plausible.
• "Offline regeneration must use the inputs captured at the time" can become my report-archive requirement today. My lab-report OCR → AI → App import pipeline has always persisted every interaction to support self-evolving learning, but if regenerating a historical report uses today's model and today's knowledge base, that report is no longer the one the user saw. Archiving the input packet as a whole is worth adding right now.
• "LLM proposals remain advisory; accepted data mappings and transforms must be reviewable and deterministic" is exactly the boundary in my GRPM report. In the mapping from SNP to supplement recommendation, which parts are rules (auditable, reproducible) and which are model-generated (advisory only) must be separated, or users cannot tell which conclusions can be traced. This is also a compliance requirement.
• The three-way label contract (pass / fail / cannot-determine) has appeared again. Last issue it was aureproof's inconclusive gate; this issue it is a three-way label contract plus agreement checks across judge cards. "Insufficient evidence" is becoming the clinical AI default rather than each project's own invention.
• On licensing: this repository has no LICENSE file. The methodology can be read; the code cannot be moved across as-is.

2. Category roundup

① AI Coding / programming agents and the skill ecosystem

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

Of the 134 entries today, the most notable thing is not which repository gained how many stars, but that "evidence" and "licensing" are sinking from "kind advice in the documentation" into "hard constraints in the code". This is happening across all four directions at once.

• First, licensing is going from "one field" to "three fields plus an audit". ClawBio/ClawBio splits each skill's licensing into license / model_license / data_license, takes the strictest, and states plainly that "empty means 'none declared', never 'permissive'"; it audited all 96 skills and found 92 MIT, 1 Apache-2.0, 1 GPL-3.0, 1 PROPRIETARY and 1 undeclared. "The code is MIT" does not mean "this project is safe to commercialise" — skills, model weights and reference data can each carry their own terms. For a one-person company building a product, this layer has to be checked before any work begins.
• Second, "conclusion strength must not exceed evidence strength" now has tooling. grotyx/Academic_writing_c_claudecode enforces "explicit association caps a sentence at associative" — when the evidence only supports association, the wording may not escalate to causation — together with retraction / concern / erratum checks, DOI consistency checks, and permission to declare an evidence gap with "None". This is where content compliance meets research integrity, and it is the ruler my evidence-bound public accounts have been missing.
• Third, the consensus in clinical AI evaluation is shifting from "emit a score" to "keep an entire reviewable evidence packet". pmanko/clinical-ai-validation-harness states outright that "Aggregate scores alone are not clinical validation", requires every run to drop run_manifest.json + events.jsonl + results.jsonl + traces, and records unknown revisions as missing rather than substituting workspace pins; athril/agentic-target-evidence scores a target through six independent lenses (genetics, biology, safety, clinical, commercial, regulatory) into a provenanced dossier, and its README says plainly that "Every verdict is LLM-generated over retrieved evidence — a preliminary research aid, not ground truth". Two independent projects, one shared subtext: multiple perspectives plus an explicit disclaimer are more honest than a finer-grained single score.
• Fourth, health products are splitting into a "full-stack OS" track and a "single-purpose MCP" track — and the defaults on both sides have changed. zilet/cairn takes the full-stack route: self-hosted, one SQLite file, PWA + REST + MCP, a coaching team that reads everything and then offers one thing, no scores, no nudges, and one-tap undo on every change that lands on its own; davidmosiah/wellness-nourish, Aryan-Jhaveri/mcp-foodguidecanada, deadletterq/mcp-opennutrition and StefanOOEOrg/mcp-food take the single-purpose route, wrapping one data source as an MCP for agents to call. Notably, the default on both tracks is now "local-first, data never leaves the machine, HTTP bound to 127.0.0.1" — half of wellness-nourish's 13 commits this issue are security hardening (restricting unauthenticated HTTP to loopback, rejecting pre-existing profile symlinks before migration, serialising and bounding shared profile persistence, hardening MCP input budgets), and foodguidecanada's 44 commits include a read-only, non-root HTTP smoke test across 24 tools. "Local-first" is no longer a personal preference — it is this product category's default architectural assumption.
• Fifth, the agent "tool layer" is being redesigned as "something that behaves like an IDE", not just a handful of added APIs. cortexkit/aft works by hoisting the host's existing read / write / edit / bash tool slots and replacing what sits underneath with tree-sitter parsing, indexed search, output compression and symbol-level operations, then layering an aft_ family of tools on top; it comes with a per-file undo stack plus named checkpoints, and with output contracts such as "when a list is truncated it must say so — shown N of M". In the same period fallow-rs/fallow (21 commits), repowise-dev/repowise (10) and DeusData/codebase-memory-mcp (17) are all doing codebase health and structured retrieval. "Let the agent see the structure before it acts" is now consensus, and "changes can be taken back" is moving from a nice-to-have to a baseline.

4. Repo index

Featured

• ClawBio/ClawBio — https://github.com/ClawBio/ClawBio
• cortexkit/aft — https://github.com/cortexkit/aft
• zilet/cairn — https://github.com/zilet/cairn
• grotyx/Academic_writing_c_claudecode — https://github.com/grotyx/Academic_writing_c_claudecode
• pmanko/clinical-ai-validation-harness — https://github.com/pmanko/clinical-ai-validation-harness

① AI Coding / programming agents and the skill ecosystem (25)

• garrytan/gstack — https://github.com/garrytan/gstack
• JuliusBrussee/caveman — https://github.com/JuliusBrussee/caveman
• Fission-AI/OpenSpec — https://github.com/Fission-AI/OpenSpec
• cline/cline — https://github.com/cline/cline
• shanraisshan/claude-code-best-practice — https://github.com/shanraisshan/claude-code-best-practice
• cathrynlavery/diagram-design — https://github.com/cathrynlavery/diagram-design
• Imbad0202/academic-research-skills — https://github.com/Imbad0202/academic-research-skills
• ChromeDevTools/chrome-devtools-mcp — https://github.com/ChromeDevTools/chrome-devtools-mcp
• alibaba/open-code-review — https://github.com/alibaba/open-code-review
• github/awesome-copilot — https://github.com/github/awesome-copilot
• VoltAgent/awesome-agent-skills — https://github.com/VoltAgent/awesome-agent-skills
• OthmanAdi/planning-with-files — https://github.com/OthmanAdi/planning-with-files
• google/skills — https://github.com/google/skills
• wanshuiyin/Auto-claude-code-research-in-sleep — https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep
• lsdefine/GenericAgent — https://github.com/lsdefine/GenericAgent
• openclaw/clawhub — https://github.com/openclaw/clawhub
• repowise-dev/repowise — https://github.com/repowise-dev/repowise
• refly-ai/refly — https://github.com/refly-ai/refly
• Devin-AXIS/iPolloWork — https://github.com/Devin-AXIS/iPolloWork
• fallow-rs/fallow — https://github.com/fallow-rs/fallow
• mohitagw15856/pm-claude-skills — https://github.com/mohitagw15856/pm-claude-skills
• ClawBio/ClawBio — https://github.com/ClawBio/ClawBio
• cortexkit/aft — https://github.com/cortexkit/aft
• mojoapp-ai/agent-skills — https://github.com/mojoapp-ai/agent-skills
• npbuilds/skill-library — https://github.com/npbuilds/skill-library

② Medical / clinical research AI and agent skill libraries (35)

• ivmartel/dwv — https://github.com/ivmartel/dwv
• afairgiant/MediKeep — https://github.com/afairgiant/MediKeep
• theislab/ehrapy — https://github.com/theislab/ehrapy
• pymedphys/pymedphys — https://github.com/pymedphys/pymedphys
• PyTomography/PyTomography — https://github.com/PyTomography/PyTomography
• xxPcy/PharmaLLM-Agent — https://github.com/xxPcy/PharmaLLM-Agent
• aphp/edsnlp — https://github.com/aphp/edsnlp
• Project-MONAI/monai-physio — https://github.com/Project-MONAI/monai-physio
• DouglasNeuroInformatics/OpenDataCapture — https://github.com/DouglasNeuroInformatics/OpenDataCapture
• imgag/megSAP — https://github.com/imgag/megSAP
• CogStack/cogstack-nlp — https://github.com/CogStack/cogstack-nlp
• rfhb/ctrdata — https://github.com/rfhb/ctrdata
• opensafely-core/opencodelists — https://github.com/opensafely-core/opencodelists
• aphp/Cohort360-FrontEnd — https://github.com/aphp/Cohort360-FrontEnd
• RASAAS/docmcp-knowledge — https://github.com/RASAAS/docmcp-knowledge
• anshk1234/MEDUCID — https://github.com/anshk1234/MEDUCID
• alanphys/LinaQA — https://github.com/alanphys/LinaQA
• grotyx/Academic_writing_c_claudecode — https://github.com/grotyx/Academic_writing_c_claudecode
• HL7/cqf-recommendations — https://github.com/HL7/cqf-recommendations
• SidneyBissoli/medical-terminologies-mcp — https://github.com/SidneyBissoli/medical-terminologies-mcp
• tvs-sde/oxford-omop-data-mapper — https://github.com/tvs-sde/oxford-omop-data-mapper
• Gilead-Public/qcthat — https://github.com/Gilead-Public/qcthat
• ma3u/MinimumViableHealthDataspacev2 — https://github.com/ma3u/MinimumViableHealthDataspacev2
• airslie/renalware — https://github.com/airslie/renalware
• cqframework/cql-tests — https://github.com/cqframework/cql-tests
• smart-on-fhir/cumulus-library — https://github.com/smart-on-fhir/cumulus-library
• pacharanero/clincalc — https://github.com/pacharanero/clincalc
• SudarshanSingh1/DR-Screening-System — https://github.com/SudarshanSingh1/DR-Screening-System
• Yash-Lalwani/TrialWright — https://github.com/Yash-Lalwani/TrialWright
• pmanko/clinical-ai-validation-harness — https://github.com/pmanko/clinical-ai-validation-harness
• broadinstitute/via — https://github.com/broadinstitute/via
• wth-lg/clinical-agent-env — https://github.com/wth-lg/clinical-agent-env
• athril/agentic-target-evidence — https://github.com/athril/agentic-target-evidence
• pkotecha-eng/faro-clinical-agent — https://github.com/pkotecha-eng/faro-clinical-agent
• SanjanaChoudhury1508/TrialLens — https://github.com/SanjanaChoudhury1508/TrialLens

③ Health management / nutrition and food AI (32)

• Snouzy/workout-cool — https://github.com/Snouzy/workout-cool
• Cawlumm/lyftr — https://github.com/Cawlumm/lyftr
• deadletterq/mcp-opennutrition — https://github.com/deadletterq/mcp-opennutrition
• rwestergren/cronometer-api-mcp — https://github.com/rwestergren/cronometer-api-mcp
• LowCarbCheck/openplate — https://github.com/LowCarbCheck/openplate
• haschildrensfoundation/pravin-elite-fitcore — https://github.com/haschildrensfoundation/pravin-elite-fitcore
• Shell3D/power-fit-hub — https://github.com/Shell3D/power-fit-hub
• ataturhan21/Omega-LC — https://github.com/ataturhan21/Omega-LC
• rfivesix/train-libre — https://github.com/rfivesix/train-libre
• davidmosiah/wellness-nourish — https://github.com/davidmosiah/wellness-nourish
• zilet/cairn — https://github.com/zilet/cairn
• healthkey-ai/promop — https://github.com/healthkey-ai/promop
• StefanOOEOrg/mcp-food — https://github.com/StefanOOEOrg/mcp-food
• StefanOOEOrg/mcp-nutrition — https://github.com/StefanOOEOrg/mcp-nutrition
• skeptrunedev/skcal — https://github.com/skeptrunedev/skcal
• jiping730/nutrition-agent — https://github.com/jiping730/nutrition-agent
• kirmalyshev/eait — https://github.com/kirmalyshev/eait
• ntsd/zerocoach — https://github.com/ntsd/zerocoach
• Synapsr/Symbiose — https://github.com/Synapsr/Symbiose
• Aryan-Jhaveri/mcp-foodguidecanada — https://github.com/Aryan-Jhaveri/mcp-foodguidecanada
• ak111pro/ha-fddb — https://github.com/ak111pro/ha-fddb
• weante/ashkb — https://github.com/weante/ashkb
• sealambda/chironium — https://github.com/sealambda/chironium
• kadato/Dietinator — https://github.com/kadato/Dietinator
• kuhyx/diet-guard — https://github.com/kuhyx/diet-guard
• smcgarrigle/ihavethereceipts — https://github.com/smcgarrigle/ihavethereceipts
• RCM7/ask.recipes — https://github.com/RCM7/ask.recipes
• 20212661/diet-agent — https://github.com/20212661/diet-agent
• tenderzzz/body-journal — https://github.com/tenderzzz/body-journal
• ElijahZhao/MetaNutri---AI- — https://github.com/ElijahZhao/MetaNutri---AI-
• arjav1181/Darkbite-Tracker — https://github.com/arjav1181/Darkbite-Tracker
• harsaas/Hybrid-Copilot-Clinical-PBM-Decision-Support-System-OKF — https://github.com/harsaas/Hybrid-Copilot-Clinical-PBM-Decision-Support-System-OKF

④ Model foundation and agent infrastructure (42)

• ollama/ollama — https://github.com/ollama/ollama
• langgenius/dify — https://github.com/langgenius/dify
• langflow-ai/langflow — https://github.com/langflow-ai/langflow
• ggml-org/llama.cpp — https://github.com/ggml-org/llama.cpp
• infiniflow/ragflow — https://github.com/infiniflow/ragflow
• OpenHands/OpenHands — https://github.com/OpenHands/OpenHands
• lobehub/lobehub — https://github.com/lobehub/lobehub
• microsoft/ai-agents-for-beginners — https://github.com/microsoft/ai-agents-for-beginners
• upstash/context7 — https://github.com/upstash/context7
• BerriAI/litellm — https://github.com/BerriAI/litellm
• pathwaycom/serviette — https://github.com/pathwaycom/serviette
• vectorize-io/hindsight — https://github.com/vectorize-io/hindsight
• Ebony-Vinyl/dsh-our-free-model — https://github.com/Ebony-Vinyl/dsh-our-free-model
• HKUDS/nanobot — https://github.com/HKUDS/nanobot
• DeusData/codebase-memory-mcp — https://github.com/DeusData/codebase-memory-mcp
• langchain-ai/langgraph — https://github.com/langchain-ai/langgraph
• agentscope-ai/QwenPaw — https://github.com/agentscope-ai/QwenPaw
• microsoft/graphrag — https://github.com/microsoft/graphrag
• Tencent/WeKnora — https://github.com/Tencent/WeKnora
• agentscope-ai/agentscope — https://github.com/agentscope-ai/agentscope
• github/github-mcp-server — https://github.com/github/github-mcp-server
• SigNoz/signoz — https://github.com/SigNoz/signoz
• simstudioai/sim — https://github.com/simstudioai/sim
• antirez/ds4 — https://github.com/antirez/ds4
• jundot/omlx — https://github.com/jundot/omlx
• comet-ml/opik — https://github.com/comet-ml/opik
• arc53/DocsGPT — https://github.com/arc53/DocsGPT
• FlashML-org/FreeToken — https://github.com/FlashML-org/FreeToken
• NVIDIA/TensorRT-LLM — https://github.com/NVIDIA/TensorRT-LLM
• NevaMind-AI/memU — https://github.com/NevaMind-AI/memU
• LMCache/LMCache — https://github.com/LMCache/LMCache
• MemTensor/MemOS — https://github.com/MemTensor/MemOS
• mcp-use/mcp-use — https://github.com/mcp-use/mcp-use
• EricLBuehler/mistral.rs — https://github.com/EricLBuehler/mistral.rs
• vllm-project/vllm-omni — https://github.com/vllm-project/vllm-omni
• kvcache-ai/Mooncake — https://github.com/kvcache-ai/Mooncake
• google-ai-edge/LiteRT-LM — https://github.com/google-ai-edge/LiteRT-LM
• Trusted-AI/adversarial-robustness-toolbox — https://github.com/Trusted-AI/adversarial-robustness-toolbox
• NVIDIA/Model-Optimizer — https://github.com/NVIDIA/Model-Optimizer
• vllm-project/aibrix — https://github.com/vllm-project/aibrix
• llm-d/llm-d — https://github.com/llm-d/llm-d
• durable-workflow/waterline — https://github.com/durable-workflow/waterline