Everything you shared on June 29, 2026, in one place. Relevance is rated against your actual setup: W5 agents + Claude Code on the Mac Mini, the Obsidian vault, and RevParPro on Supabase. Star counts are from the GitHub API fetched today; issue and PR counts marked "(video)" are read off the TikTok screen recordings.
High relevance, worth acting onMedium, mine for ideasLow, interesting but not for you now
jcodeCoding harness7,878 stars · Rust · MIT · from a GitHub link
A from-scratch coding agent harness in Rust, pitched as a faster, leaner alternative to Claude Code. Its standout is a passive, human-like agent memory system.
Why it matters to you
Auto-embeds every turn into a memory graph and retrieves by similarity, no manual store or recall calls
Background extraction plus ambient consolidation (your nightly "dream" cycle, but automatic)
Swarm mode: multiple agents in one repo with conflict notifications, mirrors your Woz/SWARM model
The catch
Young (v0.9-dev); benchmarks are self-reported on one machine
Switching means rebuilding your whole Claude Code skill/hook/agent layer
RAM and speed wins only show at 10+ concurrent sessions
VerdictStudy the passive memory architecture, do not switch tools. It is the one thing W5 does not have: retrieval that never depends on an agent remembering to check memory.
obsidian-skillsClaude Code skills38,820 stars · MIT · by kepano (Obsidian CEO) · from @chase_ai_
Five agent skills that teach Claude Code to work with Obsidian: markdown, CLI, Bases, Canvas, and a web-to-markdown extractor.
Fits your vault
obsidian-markdown: wikilinks, callouts, frontmatter, the exact syntax your vault uses
defuddle: clean web-to-markdown (you already have browser-harness for this)
Conflicts with your rules
obsidian-cli needs a running Obsidian desktop; your Claude Code is headless over SSH
Bases (.base) and Canvas (.canvas) break your ".md files only" hard rule
The markdown skill teaches image and PDF embeds your vault forbids
DecidedYou chose learn from it, do not install. I already follow your TAXONOMY rules, so no new skill file goes in. Open question still on the table: whether .base earns an exception so I can build live dashboards inside Obsidian.
SupermemoryAI memory27,922 stars · TypeScript · MIT · from @howtowebdev
An open-source "AI second brain": ingests your digital history into one memory graph, plus a single API and an MCP server to give any agent persistent memory.
Why it matters to you
The MCP server plus single memory API overlap squarely with W5's memory layer
Connectors for Drive, Gmail, Notion, GitHub with real-time sync
No vector DB config, no embedding pipelines to manage yourself
The catch
W5 memory is custom and wired into your Docker/Postgres infra today
Adopting it is a build-vs-buy call, not a drop-in
VerdictThe clearest build-vs-buy candidate for the agent memory layer. Worth a real evaluation against the homegrown W5 system before we invest more in custom memory tooling.
A curated collection of Claude Code configs (agents, skills, hooks, commands, rules) from an Anthropic hackathon winner, evolved over 10+ months of daily use. Installable as a plugin.
Patterns worth lifting
Hooks that auto-save and load context across sessions (memory persistence)
Continuous-learning: auto-extract patterns from sessions into reusable skills
Verification loops with grader types and pass@k eval metrics
The catch
Overlaps heavily with your existing skill and hook stack
No clear license on the repo; the real value is in the linked guides, not the raw code
VerdictRead it for the memory-persistence hooks and continuous-learning pattern. Same family as your setup, so cherry-pick ideas rather than installing the plugin over W5.
vibecode-pro-max-kitAgent harness1,006 stars · JS · MIT · v3.2.5 · from a GitHub link
A drop-in plan-first development harness for AI coding tools: 15 agents, 33 skills, 10 hooks, and a gated workflow that stops the agent jumping straight to code.
Patterns worth lifting
RIPER-5: seven gated phases (Research through Update-Process) before any code
/goal token: keeps a run going across sessions without losing its place
Self-healing plan-check and test-check loops
The catch
Installing collides with your existing .claude harness (reserves the vc- namespace)
"0 human gate" autopilot runs against your own verify-before-done guardrails
VerdictThe RIPER-5 gating and cross-session /goal resume are the takeaways. More formal than your ad-hoc plan-then-ship flow. Do not install over W5.
Low relevance, interesting but not for you now
SurrealDBDatabase32,531 stars · Rust · ~606 open issues (video) · from @howtowebdev
A scalable multi-model database in Rust: graph, document, relational, vector, and more in one engine, queried with a SQL-like language. Used by NVIDIA, Samsung, Tencent (per the video).
VerdictRevParPro on Supabase Postgres is plenty for hotel data. The only hook for you is its agent-memory angle (unified vector, graph, doc), which overlaps with the memory theme above but is a heavy swap. Park it.
LogtoAuth12,360 stars · TypeScript · MPL-2.0 · from @howtowebdev
An open-source Auth0 alternative: full user auth out of the box with OIDC, OAuth 2.1, SAML, RBAC, MFA, social logins, and an MCP-ready angle for AI apps.
VerdictRevParPro already runs Supabase Auth, which covers your needs. Logto is a heavier, enterprise-grade swap (SSO and RBAC). No reason to change unless you hit a wall with Supabase.
OdysseusAI workspace79,179 stars · Python · AGPL-3.0 · by PewDiePie · from @howtowebdev
A free, self-hosted, private AI workspace: run open-source LLMs on your own hardware, with chat, coding agents, email, and research in one interface.
VerdictYou already have a private agent workspace (W5) running on the Mac Mini against the Claude API. This only matters if you wanted to move W5 onto local open-source models for privacy or cost, a real tradeoff (cheaper and private vs weaker models plus a hardware bill). Watch the category, do not chase it.
If you want to act, here is the short list
Memory layer decision. Two of today's tools (jcode's passive memory, Supermemory's API and MCP) point at the same gap: W5 memory is manual and batch-consolidated. Worth one focused evaluation: keep building custom, adopt Supermemory, or borrow jcode's passive-retrieval design.
Obsidian .base exception. Still your call. Allowing .base files would let me build live deal and property dashboards inside the vault instead of separate HTML pages. The rest of the obsidian-skills decision is closed (learn, do not install).
Harness patterns to borrow. From vibecode-pro-max-kit and everything-claude-code: RIPER-5 phase gating, the cross-session resume token, and auto-save/load context hooks. These are ideas to fold into W5, not tools to install.