Useful things worth opening.

Guardrails for Pi built on pi-typesafe — "steer the agent instead of interrupting you: Jev judges." Part of a small Pi-safety ecosystem from the same author as the next entry.

TypeSafe decisions for Pi — a batched evaluation tool, terminal playground and typed API for extensions. The foundation pi-warden builds on.

Jasper Dekoninck releasing new versions of the BrokenArXiv and ArXivMath benchmarks. Announcement only; useful pointer for anyone tracking eval benchmarks.

QuiverAI's Arrow 2 — their latest models for generating precise, editable vector graphics. Announcement only; a notable data point in AI vector-graphics generation.

MiniCPM5 2B PCD weights on Hugging Face — the same model family as the 16 Sept in-browser Pi announcement. The model behind that WebGPU coding-agent demo.

An inference provider that lets Pi (a terminal agent harness) run on Claude Code Pro/Max subscriptions instead of API keys. Kenn is using a fork and admits he "just closed my eyes and let the agent figure it out" — expect rough edges.

Lets you route Pi's model calls through a Claude subscription ("allow sub usage in pi"). Same cost-saving idea as pi-claude-agent-sdk: agentic coding without per-token API billing.

Cross-vendor 3D Gaussian Splatting trainer — video to splat to mesh, running on Vulkan or CUDA. Notable for not being CUDA-locked.

A dedicated browser for responsive web development — preview a site at multiple screen sizes side by side. Built for faster, more precise responsive work.

A Rege Lab page with six little Grokbot character expressions to play with, plus downloadable editable CoAnimato files. A creative asset resource for bot and character design.

Official Claude guide to prompting Opus 5.5, steering long agent runs, and checking results in Claude apps and Claude Code. A practical reference for anyone driving long agent sessions.

A minimal bot-icon studio — browse bot icon examples and copy a prompt to generate an icon of your own. Quick asset generation for bot projects.

Cloudflare Quick Tunnels — free encrypted tunnels that expose localhost as a public URL. Cata plans to use it instead of the VS Code tunnel.

Symbolic regression implemented in Bend. A research-flavoured repo for anyone exploring the Bend language.

A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings. A drop-in skill for agentic security reviews.

Cloudflare engineering deep-dive on saving 100TB of RAM with math (and Rust). A solid systems-engineering read.

Bend 2 — a fast programming language that blocks AI mistakes via proof checking. Install with curl from bend-lang.com.

Announcement of a free agent search/fetch API — "killed Exa, Tavily and Brave", $0 with no quotas, from Tiny_Fish and MonidHQ. The pricing claims are the poster's and unverified; this link is the announcement, not documentation.

Shared by thelaggingway with just "Another one" — context is thin, but it landed in the search-tool thread. Worth a look if you're tracking search APIs.

Microsoft's ArgusAgent — a persistent, reviewed multi-agent runtime for long-horizon research and engineering. Directly relevant to anyone running agents on long tasks.

A markdown viewer "for vibe coders" — a simple, beautiful app for reading the .md files your AI writes: specs, plans, skills, agents and more. Pairs naturally with agent-heavy workflows.

LiquidAI's LFM2.5 2.6B model tuned for RLCD, shared in the context of Jev's API and concepts. A small local model for anyone experimenting with RLCD-style agent loops.

A 1B RLCD-tuned Qwen 2.5 that kenn tried locally "to get a feel of how Jev's API and concepts might feel like" — and found it did fine in a roguelike spike. Useful as a tiny testbed model before spending on bigger runs.

Announcement of a coding agent (Pi) running entirely in the browser on a 2B model via WebGPU — MiniCPM5-2B + Pi with Transformers.js and 4-bit ONNX weights. Announcement only; no repo was linked in the post.

An experimental Rust time-series database with DuckDB SQL and tiered Parquet storage — "time settles into layers." Early-stage but interesting for anyone into time-series storage design.

Ever Gauzy — an open business management platform covering ERP, CRM, HRM, ATS and project management (gauzy.co). A full open-source suite if you ever need to stand up business tooling.

Demo of a universal syntax highlighter built on the new OpaqueRange web API (Chrome 152+, 2026) — syntax-highlighting different program fragments efficiently. Worth a look for anyone building code-rendering UI.

Jakub Krehel's essay on the invisible side of design engineering — the craft that doesn't show up in the final pixels. A thoughtful read for design-minded engineers.

A reference layer for AI coding agents — real production websites with desktop and mobile captures, "real websites worth studying." Useful as grounding material when agents need to see how good sites are actually built.

A small library of playful, physics-driven React elements — including a ceiling pull-cord with real Verlet rope physics. For interfaces that should feel alive rather than flat.

A harness claiming five Qwen3.8-27B models match Claude Fable 5 on LiveCodeBench Hard — "Fable 5 level coding." The performance claims are the poster's and unverified; scarywood75 noted the implementation is Python.

A collection of the most essential transitions for web apps, plus a skill for agents and a Refine tool. Drop-in motion craft for web projects and the agents building them.

A collection of design-engineering skills covering accessibility, motion and frontend craft. Pairs well with the transitions.dev repo above for leveling up UI output.

Oskar's announcement of "e2e", a new agent-oriented testing framework — he ported React Native Pager View's test suite from Maestro to it. Announcement only; the framework itself is still being built.

A collection of beautifully designed, accessible and customizable PDF components built on Takumi. Handy if you ever need to render polished PDFs from a web stack.

An open-source 3D architectural editor with a local CLI, MCP tools and practical workflows for humans and agents. Niche but complete — 3D editing with agent integration baked in.

A professional vector design tool "where you and your AI agents work directly on the canvas." Agent-native design tooling is still rare, which makes this one notable.

Compares the real cost of AI — subscription allowances, model benchmarks and the price–performance frontier, with transparent sources. Useful before committing spend to any provider.

A pi extension giving the agent exactly one tool — doing file read/edit work in arm64 assembly — to stop Astra shelling out to python3, ruby and perl for file ops. Announcement; the actual link was promised in a follow-up tweet.

Announcement of Minara Harness — pitched as "general financial intelligence for everyone", turning ideas into signals into execution, plus coding tasks like any other agent. The capability claims are the poster's and unverified.

Daniel's post that a new herdrdev update is out with fleet management live. A product-update announcement; the herdr blog below has the substance.

A step-by-step guide to making photorealistic 3D worlds and games — reference photos, Blender assets, sub-agents, blind critics, Three.js vs Unreal Engine, and a starter prompt. The "GPT-6 Astra" framing is the site's own marketing.

An open-source Palantir alternative — track 10K+ aircraft, 2K satellites and worldwide CCTV on a 3D globe, and run Nmap, DNS, WHOIS and threat-intel scans from the browser across 20+ live feeds. Free.

An AI-native design editor pitched as an open-source Figma alternative. Worth watching if design tooling is on your radar.

The Herdr blog on 0.9, which brings local and remote machines into one TUI client, with Herdr Cloud coming next. Herdr's pitch: run agents anywhere, leave them running.

An agent-harness performance optimization system — skills, instincts, memory, security and research-first development for Claude Code, Codex, Opencode, Cursor and beyond. Directly relevant if you're tuning agent setups.

Matt Dancho's post on Microsoft's MarkItDown — turns almost any file (PDFs, Word, PowerPoint, Excel, images) into clean Markdown; claimed 100,000+ GitHub stars. Announcement only; the repo wasn't directly linked.

Google's Magika — detects the real content of any file even when disguised (malware posing as PDF, scripts hidden as images); free and open-source. Announcement only; the repo wasn't directly linked.

Tencent's "Make Every Team AI Native" CLI. The description is thin, but a Tencent-backed team-AI CLI is worth a look.

Cata's first pi extension — an ASCII showcase/mockup gallery of UIs for fast iteration. The shared link pointed at the commits page; normalized to the repo root.

Git-native persistent memory for AI coding agents — implements Google OKF v0.2 with sub-300µs in-memory BM25 search and an embedded MCP server. Claims to slash token bloat by 80%; that figure is the repo's and unverified.

Microsoft's tgrep — grep for large codebases with client/server support, claimed 7–50x faster than ripgrep and ugrep. Announcement only; the speedup claim is the poster's.

"BEAM on device" — write mobile apps in Elixir with the whole BEAM running on the device, rendering native iOS and Android UI; your coding agent attaches to the running app with real state and hot-pushed fixes. Ambitious and unusual.

"The ripgrep of AI context" — a zero-dependency C++23 CLI plus MCP server that gives coding agents a ranked, deterministic map of any repo: blast radius, tests-to-run and quality deltas. From Red Hat's emerging-tech group.

OpenAI's developer blog on using Astra to design a house, refine everyday details, and explore it through Blender camera tours and an Unreal Engine 5 walkthrough. A concrete agentic-design case study.

SponsorBar — puts a small ad in your menu bar and shares the revenue with you ("make money from your menu bar for doing nothing"), with a 10%-extra launch promo. A real launched product, though a niche one.

Browse Factory-generated wikis for popular open-source repositories — a preview of what living documentation looks like on real codebases before running it on your own. Nice for evaluating doc-generation tooling.

A GUI for your coding agents. Description is minimal, but a visual front-end for agent harnesses is a gap worth filling.

A search MCP server from baanish, who vouched for PR #4 personally. The shared link was the PR; normalized to the repo root.

Nielsen Norman Group's study guide on the psychology behind UX. A solid reference for grounding design decisions in research rather than taste.

Copy-paste animated React components from beUI. Drop-in motion without building animation primitives from scratch.

A creative component library for Canvas, WebGL and WebGPU. For interfaces that need GPU-level visuals rather than DOM elements.

A tool catalog for agents — 2,630 endpoints across 47 providers (SEO, social, enrichment, ads data) callable with one key, priced per call, plus support for your own keys and skills. Tom's 2 Sep share carried his referral param; normalized to the root domain.

Jason Zhou's announcement of "Claude for people-search" at a claimed $0.0089/lead versus $600 subscriptions. The pricing claims are the poster's and unverified; the tweet mentions a git repo "below" that wasn't captured.

Paper — "design, share, ship." A design tool baanish was scouting; no verdict from the channel yet. (Rykuuun's pricing-page share from 8 Sep was dropped as context-free.)

Kenn wanted to fork Paseo for internal use cases but held off because it was AGPL — PR #3944 has now relicensed it to Apache. The shared link was the PR; normalized to the repo root. The license change is the news: it's forkable now.

asx8678's own repo, built after burning through tokens fast. Context is thin, but it reads as a token-efficiency tool from someone feeling the pain directly.

The Elixir Forum announcing DSH, the DeepSeek Harness. Announcement only; no repo or docs link was captured.

Fresh — "The Terminal IDE." A terminal-based IDE worth a look if you live in the shell.

Monospace charts and tables for MDX and React. Handy for rendering data visuals that stay crisp in markdown contexts.

Eclipse Theia — a cloud and desktop IDE framework implemented in TypeScript. The open foundation behind several well-known IDE products.

Open-source alternative to Jira, Linear, Monday and ClickUp. A self-hostable project-management option for teams avoiding SaaS lock-in.

Tencent's Hy4 preview weights in FP8 on Hugging Face — 770B total, 49B active, 1M context, open weights. For anyone with the hardware to run a frontier-scale open model.

A GLM-5.3-Flash quant on Hugging Face. Context is thin — thelaggingway's whole review was "Huh" — but it's another fast-model option to know about.

DBX — a 20 MB tool for managing 90+ databases. Remarkably small footprint for a universal database client.

An AI agent platform whose pitch is "Everything is an Agent". Announcement-level description only; details unverified since the link was not opened.

Arize AI's open-source observability platform for LLM apps — tracing, evaluation, and experimentation for agent and RAG workloads. A standard choice for watching what production agents actually do.

Prime Intellect's framework for training AI agents with reinforcement learning. Per the sharer, its gimmick is collecting training data through its CLI — worth a look if you're into RL-for-agents; specifics unverified.

A Windows automation harness from the browser-use team (the web-agent library folks). Likely for running agents against Windows apps; details unverified.

A write-up of Spec-PTC, a speculative parallel test-time compute framework for long-horizon agents. Interesting if you follow inference-scaling tricks for agentic workloads.

An open-source agent microharness with persistent agency and recursive LLMs — "of bash, by bash". Minimal by design; good for experimenting with agent loops without a heavy framework.

Maxime Heckel's deep dive into shading-based motion design techniques. One of the better-written front-end animation articles; useful if you care about motion craft.

An AI-native browser aimed at research workflows. Product claims unverified; notable as another entry in the agentic-browser wave.

An MCP (Model Context Protocol) gateway that advertises an accuracy guarantee — unusual positioning in the MCP tooling space. Claims unverified; interesting if you run MCP servers in production.

Deep link into the pi coding agent's v2 harness documentation. Worth reading if you use pi or are designing your own agent harness.

A browsable library of Three.js components and interactive shaders. Handy for dropping 3D effects into web projects without writing raw WebGL.

A terminal tool that bills itself as "the best way to run terminals". Details unverified, but worth a look if your terminal setup needs rethinking.

Opinionated Oxlint rules that reject low-evidence TypeScript/JavaScript patterns — basically a linter config against AI-slop code. Fun and practical for teams drowning in generated code.

Salesforce's announcement of Slack Code, bringing coding-agent workflows into Slack. Announcement-level; the "oh no" from the sharer suggests healthy skepticism.

A small repo that translates "Claudish" — AI-assistant jargon — into plain English. Niche but amusing; useful if you've ever stared at a model hedging for three paragraphs.

Micro-interactions for React, "built for the half-second after a click". A focused little library for the small animations that make UIs feel alive.

A durable agent framework pitched as "Next.js for agents" — build agents with one folder, with durability baked in. Product claims unverified; the durability angle is the interesting bit.

Turns your Mac activity into a private timeline for you and your AI — local activity history as agent context. Privacy-sensitive by nature; the "private" claim is unverified.

A self-described "weird, playful" desktop app for building with AI agents — "less chatting, more building". Product claims unverified.

Vercel's announcement of fx, "a tiny, open, native coding agent" from Vercel Labs, originally an internal tool. Announcement-only source; details unverified.

A collection of "skills" for designers and engineers from Emil Kowalski (of Vercel/animations fame). Likely agent-skill style playbooks; worth browsing.

A tiny interactive explainer of how your phone finds itself — satellites, signal timing, and trilateration. The good kind of technical storytelling.

A library of animated React components. Another solid option in the crowded React-animation space; compare with the other UI libraries already in the archive.

Cua's announcement of the first open-source "Computer History", in early preview. Announcement-only; interesting for computer-use agent research.

Strips multi-vendor AI provenance marks — Unicode text hygiene, statistical rewrite hooks, and C2PA handling. Legitimate hygiene tooling; obviously dual-use.

"Crafted primitives for AI-native interfaces" — a component library aimed at the AI-app UI wave. Claims unverified.

A Zig code-intelligence server and MCP toolset for AI agents, recommended firsthand by the sharer. Native-speed code intelligence as agent context is a strong pitch.

Persistent semantic memory for coding agents — an MCP server that indexes your codebase. Useful when you have many repos and context windows aren't enough.

A JavaScript runtime for the BEAM — Web APIs backed by OTP, native DOM, and a built-in TypeScript toolchain. Genuinely novel: Erlang's concurrency model meets JS.

Unsloth's GGUF quantization of Qwen3.6-27B with MTP (multi-token prediction). For local inference of the 27B Qwen.

Alibaba's announcement that Qwen3.8's open weights are out. Announcement-only; pair with the unsloth GGUF quant below for local runs.

Unsloth's GGUF quantization of Qwen3.8-27B for local inference. The practical artifact behind the open-weights announcement above.

A JetBrains IntelliJ ACP (Agent Client Protocol) adapter for the pi coding agent. Niche but exactly what you need if you want pi inside IntelliJ.

Woxi's comparison of itself against Mathematica — Woxi being an interpreter for a subset of the Wolfram Language. A decent way into what Woxi is and where it stands.

GitHub's official toolkit for spec-driven development — write the spec, let agents implement. From GitHub itself, so it carries weight in the SDD workflow trend.

An open-source, model-agnostic, BYOK GitHub bot that runs in GitHub Actions. Bring your own key, pick your model, automate your repo — the self-hosted answer to proprietary coding agents.

DeepSeek's agent harness in developer preview — "everything is a plugin", with source released alongside. Every agent capability is a swappable plugin; the plugin architecture is the interesting idea here. Claims unverified.

A multiplayer environment for coding with agents — shared threads where teammates and agents build and review together, with full conversation and code context. Product claims unverified.

The Rust Foundation's quality standard for Rust training. Reference material if you're evaluating Rust courses or building one.

A minimalist coding agent that runs natively in the terminal. The "minimalist" claim comes from the sharer's one-line description and is unverified.

Turso's introduction to AgentFS, which Rykuuun found interesting. It appears to be an agent-oriented filesystem from Turso; anything beyond the docs link is unverified.

Rykuuun shared this asking "this is pi-vcc right?", suggesting it's related to a "pi" coding agent with very large context. Details are unverified from the message text alone.

A GitHub repo named "pi", shared without description. Given the surrounding chatter about pi-vcc, likely an agent project; unverified from the message text.

A GitHub repo shared without description. What it does is unverified from the message text alone.

A design-focused site, shared link-only. Its exact purpose is unverified from the message text; the domain suggests a design inspiration or resource site.

A site by Jakub Antalik, shared link-only. The name suggests gooey/morphing visual effects; unverified from the message text.

A gallery of collected user interfaces for design inspiration. Shared link-only; the description is inferred from the site's name and unverified from the message.

An agent "skill" on GitHub. What it teaches the agent is unverified from the link-only share.

A 2020 article by Daniel Rotter arguing for CSS layouts that avoid z-index. The thesis is in the URL slug; the article's quality is unverified.

A demo site showcasing a "beautiful UI". Shared link-only; what it demonstrates is unverified.

A GitHub repo, likely a motion/animation library given the name. Shared link-only; unverified.

A personal "vault" site by Arlan Marat. Contents are unverified from the link-only share.

A design-styles reference site. Shared link-only; unverified beyond the domain.

A directory of open-source alternatives to popular software. The description is inferred from the site's name; unverified from the message.

The josh project on GitHub, shared link-only. What it does is unverified from the message text.

A design checklist site. Shared link-only; unverified beyond the domain.

asx8678's entire review was "crazy". What the repo does is unverified from the message text.

A Cloudflare blog post about Kitesurf. The post's claims are unverified; it's a vendor blog.

An agent skill themed on Linus Torvalds, presumably encoding his review style for code agents. Shared link-only; unverified.

Rykuuun found it "interesting...". What it does is unverified from the message text.

Artificial Analysis's model page for Qwen3-8 Max, with benchmarks and pricing. Shared link-only; the page's data is third-party and unverified.

A GitHub repo named opencli, shared link-only. What it does is unverified from the message text.

Shared link-only. Purpose unverified from the message text; the domain suggests a design/engineering resource.

A Liquid AI blog post about "Antidoom". The post's claims are unverified; it's a vendor blog.

A Wafer blog post benchmarking Kimi K3 on AMD MI355X hardware. Claims unverified; vendor blog.

A deployed demo called Snowflow. What it demonstrates is unverified from the link-only share.

The demo repo behind the Snowflow site above. Contents unverified from the link-only share.

A shell auto-completion tool for your terminal. Details unverified from the link-only share.

An MLX port of moonshotai's Kimi-K3 (2.78T multimodal MoE); the weights are converted from the original release and requantized. Conversion quality is unverified.

A self-hosted AI agent console ("Tu agente AI self-hosted") that thelaggingway is giving a go. The self-hosted claim is unverified beyond the site's tagline.

Permanent memory for AI agents, from Victor Taelin. Details unverified from the link-only share.

An agent-ready Sanity + Next.js kit. Details unverified from the link-only share.

A free, framework-free AI chat UI kit for AI projects. Details unverified from the link-only share.

A real-time global intelligence dashboard. Details unverified from the link-only share.

Premium React micro-transitions and interaction components. Details unverified from the link-only share.

Run the pi coding agent as a service. Details unverified from the link-only share.

Beautiful animated UI components for React. Rykuuun shared it as a front-end dev finding; details unverified beyond the site's tagline.

"Make the Web Spatial" — a playground for spatial web experiences. Details unverified from the link-only share.

Privacy-first browser tools. Details unverified from the link-only share.

A desktop app that generates 3D models from images using local AI. Details unverified from the link-only share.

Rebuilds the object in a reference image as code-only procedural Three.js. Details unverified from the link-only share.

Shared without description; the name suggests an OpenAI-compatible server with context compaction for the pi agent. Unverified.

Plasma AI's Fractal wraps Claude Code and Codex in bounded recursive loops. Details beyond the truncated preview are unverified.

Static checks for ordinary TypeScript layout code, from chenglou. Details unverified from the link-only share.

UI/UX animation inspiration for mobile and web apps. Details unverified from the link-only share.

A Cursor-style visible follow-up queue with inline editing for the Pi agent. Details unverified from the link-only share.

Use your x.com Premium/SuperGrok subscription in pi with OAuth login, auto-refresh, and live model discovery; supports grok-4.5, grok-composer-2.5-fast, grok-build, grok-4.3, and x_search. Feature claims come from the repo description and are unverified.

A vendor verifier from MoonshotAI; baanish asked if anyone has run it. Details unverified from the message text.

An agentic coding setup that Rykuuun found "more flexible" than alternatives. Details unverified from the message text.

A proxied shell + exit-node picker for Tailscale userspace networking, with egress through Mullvad or other providers. Details unverified from the link-only share.

Docker & Linux containers native to Mac — free, open source, one shared VM for all containers, claiming ~4.7x less idle memory than per-container setups. The memory claim is the vendor's and unverified.

An agent skill that generates beautiful architecture diagrams with dark/light theme toggle and PNG/JPEG export. Details unverified from the link-only share.

A fast Rust port of Steve Yegge's beads — a local-first, non-invasive issue tracker storing tasks in SQLite. Rykuuun "really loved this tool" and says it makes things easier to manage.

A loop library from ForwardFuture Signals, shared as a design reference. Details unverified from the link-only share.

Flue 1.0 Beta announcement — "the open framework for building agents without lab lock-in: write once, deploy..." Details beyond the embed preview are unverified.

"Supercharge Your LLM with the Fastest KV Cache Layer." The performance claim is the project's and unverified.

Use Claude Code, Codex, and Gemini CLI through the Agent Client Protocol (ACP) inside pi. Details unverified from the link-only share.

A fix by monotykamary — diff.print() (default Patch format) was stripping the +/-/space line-origin prefixes from output; this PR prepends them back.

A validate-then-repair extension for pi that fixes common LLM tool-call mistakes (null fields, stringified arrays, wrong field names, anchor bleed) before tools execute. Tom: "technically their tool repair brought over to pi (+ my additions)."

A fix for fff — the /fff-mode override was only stored in an in-memory variable lost on /reload and session restart; this PR persists fff-mode across reloads and session resumes.

"The fastest and the most accurate file search toolkit for AI agents, Neovim, Rust, C, and NodeJS." The performance claim is the project's and unverified.

Analyze coding (agent) CLI token usage and costs from local data.

"Write HTML. Render video. Built for agents." Details unverified from the link-only share.

Generate HD short videos with one click using AI LLMs (the repo description is in Chinese). Details unverified from the link-only share.

A local-first private diary app. Details unverified from the link-only share.

MiniMax's M3 announcement — frontier coding capability, 1M context, native multimodality, and a new MSA architecture. All vendor claims, unverified.

A Codex/ChatGPT multiple-account load balancer and proxy with usage tracking, a dashboard, and OpenCode compatibility. Tom: "gonna try first."

A CLI tool to switch and manage Codex accounts. Details unverified from the link-only share.

"Built to extend you. 10× your day" — an AI extension of you across every device and person: takes your calls, writes your emails, and more. Vendor claims, unverified.