AI Open Source Radar — July 16, 2026

Method: Trendshift daily (stars today) as the primary signal, cross-checked against findarepo’s 7-day growth. Repos whose recent growth has stalled or decayed are excluded no matter how large their star totals, as are the big projects that have been holding top ranks for months.

Nutlope/hallmark — +967 ⭐ today (6.2k total) · 2 days running

https://github.com/Nutlope/hallmark

hallmark is an anti-AI-slop design skill that stops the UIs generated by Claude Code, Cursor and Codex from looking generated. Given a brief, it first picks a macrostructure for the page, dresses it in one of twenty themes, then pushes the result through fifty-seven slop-test gates plus a pre-emit self-critique. Its central claim is structural variety rather than visual variety: feed it two different briefs and you get two different sites, not colour-swaps of the same template. It ships four operating modes and four design genres:

Design (default) — the flow for new pages
audit — scores an existing target against an anti-pattern list
redesign — re-skins inside existing implementation boundaries
study — extracts design DNA from a URL or screenshot
Genres — editorial (default, the canonical anti-slop voice), modern-minimal (Stripe / Linear / ElevenLabs school), atmospheric (Suno / Runway dark-AI-tool school), playful (post-Linear soft school)

It first hit #5 on GitHub Trending on July 13, and its daily gain widened from +853 yesterday to +967 today — accelerating for a second straight day. The appeal appears to be the explicit gating: it refuses the on-distribution defaults every LLM was trained into, rather than hoping the model avoids them.

Practical angle: when you’re mass-producing landing or campaign pages with agents and they all come out looking alike, audit mode lets you check before shipping and force per-brand differentiation.

Tags: #AntiSlopDesign #DesignSystems #LandingPages #PromptGates #ClaudeSkills #FrontendGeneration

modiqo/waggle — +596 ⭐ today (613 total) · new

https://github.com/modiqo/waggle

waggle is an MCP-native reference layer that replaces pasting whole context into agent handoffs with a single ~30-byte token. The token is not just a path: behind it sits an attribution manifest recording who minted it, for which channel, from which parent (Ed25519-signed when the host holds an identity), along with variants — different projections for different consumers. When an agent resolves the token it presents its context and a sealed, deterministic matcher returns that agent’s own view; everything after is an event in a payload-free append-only log, so you can count which subagent actually read its input without ever seeing the data. That is precisely the author’s argument — a bare filesystem path cannot answer three questions a handoff eventually faces: was it read and which parts, which version, and reachable from where. The feature set spans the full mint / resolve / record / mutate / funnel / read / search / query / map loop, content-addressed snapshots pinned at mint with windowed reads under byte budgets, consumption contracts via --require with coverage verification, a tree-sitter symbol lens (Rust, Python, TS/JS, Go), and daemon-to-daemon federation out to Cloudflare’s edge (a Durable Object per tenant, resolve p50 1.2 ms). It leads with hard numbers — multi-agent systems consume roughly 15× the tokens of a chat session, and about 37% of multi-agent failures trace to the handoff seam — and having shipped v0.5.3 just yesterday, it took 596 of its 613 total stars today, making it effectively a same-day discovery.

Practical angle: handing subagents a token instead of pasting the plan into every prompt removes the re-billing, and the funnel catches any agent that answered without opening its input.

Tags: #AgentHandoff #MCP #RustLang #ContentAddressing #ContextBudget #EventSourcing #Ed25519

oso95/scroll-world — +583 ⭐ today (887 total) · new

https://github.com/oso95/scroll-world

scroll-world is a Claude Code skill/plugin that turns any brand or industry into a scroll-scrubbed 3D world landing page. The output feels like one continuous camera flight through a generated world, with scroll position scrubbing the video’s timeline — the same technique behind Apple’s scroll-through product pages. The key detail is that the camera genuinely moves and scroll only drives time. Images come from Higgsfield (GPT Image 2) as cohesive isometric diorama scenes, camera flights from Seedance image-to-video, and the scenes are stitched with connector clips built from matching boundary frames. The workflow interviews you about brand, message, visual direction and scene order, then generates stills, scene-entry videos and connectors in sequence, finishing with a portable JS scrub engine that maps scrolling to video time, lazy-loads assets and blends the seams between clips. The repo ships documentation, prompt templates, batch generation scripts, a vanilla-JS scrub engine that drops straight into plain HTML, and optional background knockout. It first trended at #17 for JavaScript on July 8, went quiet, and has now taken 583 of its 887 total stars today — a very steep slope.

Practical angle: product launch pages and campaign microsites without outsourced motion graphics, and since the scrub engine is standalone it ports into existing static sites.

Tags: #ScrollScrubbing #GenerativeVideo #DioramaArt #LandingPages #ClaudeSkills #VanillaJS

cobusgreyling/loop-engineering — +448 ⭐ today (7.6k total, +1.3k/7d per findarepo) · new

https://github.com/cobusgreyling/loop-engineering

loop-engineering packages patterns, starters and CLI tools around one thesis: the leverage point has moved from crafting individual prompts to designing the control systems that orchestrate agents over time. The author puts it bluntly — “you shouldn’t be prompting coding agents anymore, you should be designing loops that prompt your agents” — and credits Addy Osmani and Boris Cherny (Anthropic) as starting points. The loop architecture it teaches runs Schedule → Triage skill → STATE.md → Worktree → Implementer → Verifier → MCP/PR → Human gate, and it is deliberately tool-agnostic so the same shape works in Grok, Claude Code, Codex, OpenClaw, Opencode and GitHub Actions. The tooling is:

loop-init — scaffolds your loop, prints a Loop Ready score and the first run line
loop-audit — checks an existing setup
loop-cost — tracks spend
loop-context — a stateful memory manager and circuit breaker for long runs
Seven interactive patterns — open source, production-ready, one command in any repo

Its 7-day average ran around 190 stars a day; today’s +448 is more than double that, putting it in an acceleration window.

Practical angle: drop standard human gates and circuit breakers into overnight agent pipelines to stop runaway loops and cost blowouts.

Tags: #AgentOrchestration #LoopEngineering #CircuitBreaker #CostControl #StatefulAgents #CLITooling

Today’s Summary

Today’s risers split cleanly into two infrastructure repos about how to run agents (waggle’s handoff reference layer, loop-engineering’s loop control patterns) and two design repos about making agent output not look like agent output (hallmark, scroll-world). waggle and scroll-world are effectively same-day discoveries, having taken most of their total stars today, while hallmark widened its daily gain over yesterday for a second straight day of acceleration. Yesterday’s picks — colibri, orca, Vibe-Trading and destructive_command_guard — all lost momentum and were dropped.