Veiovis

A set of agent skills for Claude Code and Codex, written in Python, that cover the life of a pull request. full-send carries a change to a merged PR with nobody watching; full-stop halts when it is ready for you, and council loops three models from different providers until they agree. Skills call one another, so the README says which to install together, and they are written to be edited.

Made by t3dotgg on GitHub, not me.

GitHub - t3dotgg/tstack: The agent skills I use every day. Works with Claude Code and Codex.

Hob is an MIT-licensed desktop app for Mac, Windows and Linux that runs open-weight models on your own machine, so chats and files never leave the system. It picks a model that fits your hardware on first launch, then reads and edits files in a folder, runs sandboxed Python for charts and conversions, and recognizes images. Web search and MCP connectors are the only parts that go online.

Made by pancomplex on HN, not me.
HN: news.ycombinator.com/item?id=50246238

GitHub - FrigadeHQ/hob: An easier way to use AI on your own computer. A free, open-source alternative to Claude and ChatGPT that runs open-weight models like Gemma and gpt-oss locally on Mac, Windows

A model read 45,047 emails from eight Enron mailboxes and built a 214-page wiki of the firms, regulators and banks around the company, with no human edits. Each of its 49,401 facts links to the message it came from, and a timeline slider shows only what the mail knew by a given date, so facts appear and lapse as the story unfolds. Don't skip the methodology: the model kept slipping in what it already knew about Enron, and they show how they caught it and what the run cost.

Made by this+that. Jeff Reynar, a co-founder, posted it to Hacker News: news.ycombinator.com/item?id=50207620

What did Enron know? The Enron Brain

Point it at a folder of Markdown (an Obsidian vault, meeting notes, an agent's memory) and it builds a local index. Rather than summarising, it drafts questions from your recent notes and lets each one pull in related passages from across the years, so an agent can trace a thread through thousands of notes. A short .enzyme program sets what it should notice. It ships as one binary under 30 MB with its embedding model inside, open source under Apache 2.0.

Made by Joshua Pham. Hacker News thread: news.ycombinator.com/item?id=50221066

GitHub - byenzyme/enzyme: Local-first compile step for knowledge bases. Save 350x cost, 1000x speed vs. frontier models

When an AI reaches a design choice, it writes the question into your repo: the options, a recommendation, its rationale, its confidence, and a falsifier, the finding that would prove it wrong. You answer in a line, from your editor or your phone, and the record stays in Git beside the code it explains. It comes as a repo template, a VS Code extension with an MCP server, and an iOS app.

Made by Jérémie Lumbroso. Hacker News thread: news.ycombinator.com/item?id=50227041

ADRs4AI — Your thinking stops disappearing. Your decisions start compounding.

One API call creates an agent, and Omnara gives it sandboxes, secrets, permissions and the usual harness tools, on the model you pick and on your own servers if you like. The design is the interesting part: every step an agent takes is appended to a log in Postgres and any free worker can run the next one, so an agent outlives crashes and redeploys and holds no worker while it waits. Apache 2.0, with a hosted version as well.

Made by the Omnara team. Kartik from Omnara posted it to Hacker News: news.ycombinator.com/item?id=50183892

GitHub - omnara-ai/omnara: The open-source managed agent platform. A self-hostable alternative to Claude Managed Agents and OpenAI's Agents API.