andrew.dunn.dev

June 2026 Links

~6 min
linkscollecting
A small child holding a purple pinwheel, an older child on a balance bike ahead, on a tree-lined path heading to the park
June 2026 · Grand Rapids, MichiganA pinwheel out front and a balance bike up ahead, on the way to the park. Photo taken in June 2026 in Grand Rapids, Michigan with an iPhone 17 Pro.

AI Tooling

  • Zed, Rust-based editor built around agentic editing as a first-class collaboration mode, not a sidebar plugin.
  • Claude Code Champion Kit, Anthropic’s playbook for engineers driving Claude Code adoption inside their team. The framing (multiplier, not help desk) is the part worth stealing for any internal-tool rollout.
  • agent-browser, Vercel Labs’ browser automation CLI built for AI agents. (See also agent-browser.dev.)
  • obscura, a headless browser for AI agents and web scraping.
  • hyperframes, HeyGen’s “write HTML, render video” pipeline, built so agents can produce video by emitting markup.
  • tessera, workspace for organizing AI coding sessions across projects, collections, tabs, panes, and Git worktrees.
  • agent-skills-eval, a test runner for agentskills.io-style AI agent skills.
  • huashu-design, HTML-native design skill for Claude Code: high-fidelity prototypes, slides, animations, 20 design philosophies, and a 5-dimension review pass, all agent-agnostic. See also the English mirror, auto-synced via a translation pipeline.
  • SkillSpector, NVIDIA’s security scanner for AI agent skills: it detects malicious patterns and vulnerabilities before a skill is installed. The natural gate for any skills registry, and the companion to agent-skills-eval above.
  • awesome-legal-skills, a curated collection of ~140 agent skills for legal work (contract review, litigation support, privacy compliance), organized by practice area with jurisdictional variants in French, German, Portuguese, and Icelandic. Worth noting less for the law than as a signal: the skills idiom is spreading into professions well outside software.
  • Ponytail, a ruleset that makes a coding agent “think like the laziest senior dev in the room.” Its decision ladder (skip the code, use a built-in, reuse a dependency, then write only what’s left) is the discipline I keep trying to encode by hand.
  • Understand Anything, turns any codebase into an interactive knowledge graph you can explore, search, and ask questions about, across several AI coding platforms.
  • Vane, a privacy-focused AI answering engine that runs entirely on your own hardware, pairing web search with local or cloud models to return cited answers.

Agent Frameworks

  • EDDI, config-driven engine that turns JSON into production-grade AI agents. Multi-agent orchestration, 12+ LLM providers, MCP and A2A protocols, RAG, persistent memory, with EU AI Act, GDPR, and HIPAA compliance baked in. Built on Quarkus.
  • gitagent-protocol, framework-agnostic, git-native standard for defining AI agents. (See gitagent.sh.)
  • agentfm-core, peer-to-peer network that turns everyday computers into a decentralized AI supercomputer, running workloads across a global mesh of idle CPUs and GPUs.

Memory and Context

A whole category showed up this month: a dedicated memory layer that sits beside the model and accumulates what it learns.

  • supermemory, a memory and context engine that runs fully locally, billed as “the Memory API for the AI era.” It extracts facts from conversations, keeps a user profile, and serves context back across sessions.
  • Hindsight, Vectorize’s agent memory system: it stores experiences, world facts, and learned mental models so an agent gets more accurate the longer it runs.
  • GBrain, the same idea pointed at personal and team knowledge, with hybrid retrieval, auto-linking, and autonomous enrichment. Its pitch (“search gives you raw pages, GBrain gives you the answer”) is the synthesized-answer-with-citations shape several of these are converging on.

Models and Inference

  • omlx, LLM inference server with continuous batching and SSD caching for Apple Silicon, managed from the macOS menu bar.
  • Talkie, a 13B language model trained exclusively on English text from before 1931, built to study how models behave when trained on historical rather than contemporary data.
  • Nous Psyche, Nous Research’s open infrastructure for training LLMs over the open internet: the DisTrO optimizer compresses coordination traffic by orders of magnitude, nodes coordinate through Solana, and the first production run pretrained a 40B model on 20T tokens across volunteer hardware, which they describe as the largest pretraining run conducted over the internet. The training-side sibling of agentfm-core above.

Research

  • Meta-Harness, reference code for Stanford IRIS Lab’s Meta-Harness paper on agent harness design.
  • Claude Science, Anthropic’s research workbench: persistent Python and R kernels, native rendering for proteins, genomic tracks, and chemical structures, and connectors to 60+ scientific databases. The piece I’d want in every analysis workflow is the background reviewer that flags citations that don’t check out and figures that don’t match the underlying code.

Self-Hosting

  • Noodle Gallery, open-source self-hosted alternative to Google Photos: AI-powered search, face recognition, shared family spaces, and cloud storage, all under your control.
  • Shared Spaces, the Noodle Gallery feature I want Immich to ship quite badly: multiple people contribute photos from their own accounts into one chronological timeline, without merging libraries.
  • Morningside, home management platform that centralizes household finances, documents, maintenance records, and projects. AI search runs over your own uploaded files, not the open web.
  • DefGuard, a self-hosted remote-access platform that rolls WireGuard VPN, identity and access management, MFA, and network access control into one. The all-in-one answer to the SSO stack I assembled by hand.
  • Checkmate, open-source self-hosted monitoring for server hardware, uptime, response times, and incidents, with the kind of visualizations most uptime tools skip.
  • PriceGhost, self-hosted price tracking that watches product prices on any site and pings you on a drop, a target price, or a restock.
  • IINA, the mpv-based macOS video player, paired with iina-jellyfin, a plugin that turns it into a proper Jellyfin client: sidebar library browsing, playback progress synced back to the server, auto-queued next episodes, and Quick Connect login.
  • docker-steam-headless, a containerized headless Steam client with Proton and GPU support, streamable over Moonlight, Steam Link, Steam Remote Play, or straight from the browser. Game streaming as just another workload in the rack.

Music and Sound

  • Banan-a-Synth, MicroKits’ pocket-sized sound machine and musical science lab: 12 instrument sounds, eight effects, four-note polyphony, and alligator clips so you can turn actual bananas (or anything conductive) into a keyboard. Already 10× over its Kickstarter goal.
  • Brand New Noise, handmade lo-fi audio instruments (voice recorders, kalimbas, toy pianos, tongue drums) designed to blur the line between toy and tool.

Reading

  • 0to1 Engineer at Mozilla.org, the job title alone is worth noting: the named archetype is shifting from “senior software engineer” to “the person who takes things from zero to one.” Roles are catching up to how teams actually describe the work.
  • Cognitive Surrender, Addy Osmani on the difference between cognitive offloading and cognitive surrender: offloading hands the machine work you could still check, surrender hands it the thinking itself, and the resulting comprehension debt compounds quietly because generated code looks right. The posture framing (thinking partner versus thinking replacement) is the useful takeaway.

Learning

  • Human-Centered Systems Thinking, IDEO U’s five-week cohort course on pairing systems thinking with human-centered design to shift complex systems. The two halves are exactly the lens I keep reaching for, and rarely see taught together.