Weekly digest
The week on HuiaHub: sandboxed agents, RAG tooling and local models
41 new HuiaHub pages this week: sandboxing coding agents, RAG retrieval, local models on Macs and phones, Claude Haiku 5.5 and six how-to guides.
Two threads run through this week's 41 new pages. Several cover keeping coding agents inside a sandbox, and a cluster on retrieval runs from picking an embedding model to searching screenshots instead of text.
New agent skills
- Claude Academy guide answers a question about using Claude first, then suggests a matching Claude Academy course or tutorial.
- Discernment nudge ends a substantive answer with two or three follow-up questions that help the user check facts and spot missing context.
- Slack GIF creator pairs Slack's size guidance with a small Python toolkit for drawing animated GIFs with Pillow.
- The py skill from pysheeet sends the agent to pythonsheets.com for Python questions and tells it to fetch pages rather than answer from memory.
- Dataframely conventions teach an agent how Quantco's dataframely validates polars data frames.
- Upstash Vector skill is documentation only: a TypeScript SDK quick start plus reference files on methods and pitfalls.
Open-source projects
- DigUp is a free macOS menu bar app that searches your files by content, using EmbeddingGemma 2 running locally.
- llmfit is a Rust terminal tool that inspects your hardware and estimates which open models will run, and how fast.
- TencentDB Agent Memory is a self-hosted Tencent Cloud service that stores what agents learn and shares it across agents and team members.
- iFixAi audits AI agents and models with 60 free inspections and returns an A to F grade with a written report.
- oMLX is a local inference server for Apple Silicon Macs with OpenAI-compatible and Anthropic-compatible endpoints and a KV cache that offloads to the SSD.
- turbovec is a vector index for Python and Rust that stores embeddings at 2 or 4 bits per dimension.
- PixelRAG is the official code for a Berkeley paper on retrieving over screenshots of web pages, PDFs and images instead of extracted text.
- qwen-audio-agent puts a realtime voice model in front of a coding agent, so longer jobs run while the conversation continues.
AI tools
- Scrape Anything is Context.dev's scraping endpoint, turning a URL into Markdown, HTML, JSON or screenshots.
- Lune is an MCP server that lets an agent search and cite computer science papers from A and A* conferences.
- KernelAI runs 53 open-weight models offline on iOS and Android phones.
- ej is an 11 MB open-weight model from HITL Systems that answers routing, yes/no and priority questions about text on a CPU.
- AgentDock shows each running coding agent as a status dot on a Mac's screen edge and collects their questions in one Inbox.
- PocketWebTools for Mac runs open models up to 35B parameters offline on Apple Silicon for chat, transcription, OCR and more.
- Refs turns launch films and explainers into blueprints and prompt recipes that Claude and other agents can build from.
- ngram 2.0 makes onboarding, training and demo videos from a prompt or a screen recording.
Videos
- Cerebras and Gimlet Labs on inference speed is a 22-minute talk on paged KV cache, chunked prefill and system-level disaggregation.
- Robert Brennan of OpenHands compares four ways to sandbox a coding agent in 14 minutes.
- Fireworks AI's train-to-deploy talk covers training open models on your own data and redeploying them in a continuous loop.
- Three questions for agent engineers is a LangChain talk on finding, reproducing and fixing agent failures from production traces.
- freeCodeCamp.org's n8n course goes from a first workflow to agents, Slack approvals and error handling in 43 minutes.
- IBM Technology's distributed inference explainer covers data, pipeline, tensor and expert parallelism in about nine minutes.
- Building a GLM-5.3-Flash-style model in Python is a 44-minute walkthrough with all code on GitHub.
- Weaviate Podcast #146 talks with Siddharth Gollapudi about using an LLM's attention over a million-token context as the retriever.
Guides
Six how-to guides went up this week.
- Sandboxing AI coding agents covers setup in Claude Code, Codex, Copilot and Docker, and the gaps each boundary leaves open.
- Choosing an embedding model for RAG weighs language, vector size, hosting and price, then has you test on your own documents.
- Evaluating a RAG pipeline with Ragas builds a small test set, runs a first score and compares two versions.
- Prompt caching across OpenAI, Claude and Gemini compares each API's rules and shows how to order prompts so the cache hits.
- Testing an MCP server works in three layers: the MCP Inspector, unit tests and a scripted client check in CI.
- LoRA fine-tuning on one GPU prepares data, trains with PEFT and TRL or Unsloth, and compares against the base model.
Five news write-ups sit alongside them.
- Claude Haiku 5.5 cuts Haiku input pricing to $0.10 per million tokens and adds effort control.
- GitHub's secret detection model moves AI secret detection to a fine-tuned model billed in AI Credits.
- Copilot local sandboxing is now generally available and limits what agent-run commands can touch.
- Agent Lightning v1.0 is Microsoft's rebuilt 3,500-line RL framework that trains agents through the harnesses they already run.
- NVIDIA's Nemotron write-up explains how SFT, RL and test-time loops took Nemotron 3 to gold-level IOI and IMO 2026 results.
Sources
- 1.github.com/microsoft/agent-lightningRepositoryPrimary
- 2.One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMORepositoryPrimaryhuggingface.co
- 3.AlexsJones/llmfitRepositoryPrimarygithub.com
Spotted an error? See our corrections policy.
Enjoyed this? Get the weekly brief.
New guides, Project & Tool of the Week and the AI developments that matter. One email, no spam.
Read next
ExplainerAgents & automationAI agents found in US bank accounts: what banks are blocking and why
See which banks are restricting AI agents in bank accounts, how read-only access differs from credential sharing, and which claims remain unverified.
10 min read
GuideCoding & dev toolsPHP critical security update CVE-2026-109618: Versions and patch steps
Check the PHP critical security update CVE-2026-109618, identify affected servers, install fixed PHP versions, and apply temporary settings if needed.
12 min read
NewsHow NVIDIA fine-tuned Nemotron 3 for gold-level IOI and IMO 2026 results
NVIDIA details how SFT, RL and test-time loops took Nemotron 3 to gold-level IOI and IMO 2026 results, and which checkpoints and datasets are open.
4 min read
Related open-source projects
TencentDB Agent Memory is a self-hosted memory service from Tencent Cloud that stores what AI agents learn and shares it across agents and team members. It turns past work into four kinds of memory asset. Chat Memory keeps preferences, facts and decisions, distilled from raw conversation through layers named L0 Conversation, L1 Atom, L2 Scenario and L3 Persona. Skills capture reusable procedures with versions and validation rules. LLM-Wiki builds linked pages from documents, and CodeGraph indexes symbols, files and call relationships so an agent can check what a change might affect. A web panel called Memory Hub tracks ownership, versions and access, with private, team and ACL-based visibility. Agents connect through a proxy. You point the agent's base URL at it, and the README says no plugin, hook or MCP server is needed. Listed clients include Claude Code, Codex, CodeBuddy, WorkBuddy, Hermes and OpenClaw. Setup starts three services (memory-core, memory-hub and proxy) with Docker and a start script, and asks for two sets of LLM settings. The README reports a PersonaMem score of 76% with the system against 48% without it, which is the project's own figure. Team Memory is labeled beta.
- Stars
- 28K
- Language
- TypeScript
- Momentum
- Established
qwen-audio-agent is a realtime voice runtime that sits in front of a coding or task agent. You talk to a voice model, longer jobs go to a backend agent while the conversation continues, and when a task finishes its result comes back into the same conversation. The voice side is swappable. Cloud options listed in the README include Qwen Audio 3.0 Realtime, GPT-Live / OpenAI Realtime, Google Gemini Live and Doubao Seeduplex 3.0 Realtime, each needing that provider's API key. Hugging Face Speech-to-Speech and MiniCPM-o 4.5 can run locally. Backend agents connect mostly over ACP: the list covers Qwen Code, OpenCode, OpenClaw and Kimi Code, with Codex and Claude Code joining through an external ACP adapter. You can use it from a WebUI, a terminal TUI or a desktop floating orb on macOS, Windows and Linux. The README also lists parallel background tasks you can query or cancel, cross-session memory, and scheduled reminders. The package is published on npm under Apache-2.0, a technical report is on arXiv, and the repository sits in the QwenAudio GitHub organization.
- Stars
- 2.9K
- Language
- JavaScript
- License
- Apache-2.0
- Momentum
- Emerging
PixelRAG is a retrieval system that turns web pages, PDFs and images into screenshot tiles and searches those images instead of extracted text. It is the official code for the arXiv paper "PIXELRAG: Web Screenshots Beat Text for Retrieval-Augmented Generation", written by researchers at Berkeley SkyLab, BAIR and Berkeley NLP. The pipeline runs in stages. pixelshot renders a document into tiles with a headless Chromium browser driven over CDP. The pixelrag command then chunks and embeds the tiles with Qwen3-VL-Embedding and LoRA adapters fine-tuned on screenshot data, builds a FAISS index, and serves search through a FastAPI app on CPU or GPU. Each stage is its own pip extra, so a project installs only the parts it uses. The team also runs a hosted index of 8.28M Wikipedia pages behind a live API, which takes text or image queries; the README says it needs no API key. The renderer doubles as a Claude Code plugin called pixelbrowse, which has Claude screenshot a page and read the image, and as an npm package for opencode. Training code sits in a separate train/ project with its own pinned PyTorch and transformers versions.
- Stars
- 10K
- Language
- Python
- License
- Apache-2.0
- Momentum
- Established
turbovec is an in-process vector index for Python and Rust that stores embeddings at 2 or 4 bits per dimension and searches the compressed vectors directly. It is written in Rust with Python bindings and uses TurboQuant, a quantization method from Google Research. The README's headline figures: ten million 768-dimensional embeddings that need 31 GB as float32 fit in 4 GB, and search runs 4.4× faster than FAISS FastScan at 4-bit, at matching recall. Its benchmark table puts the 4-bit speedup at 3.8–5.1× across eight test cells on GCP ARM and x86 machines. These are the project's own benchmarks, and the scripts and raw JSON are published alongside them. No training step is needed, so a vector is searchable as soon as it is added, and removal takes microseconds. sync() writes only what changed since the last call. IdMapIndex keys vectors by your own uint64 ids and accepts an allowlist filter at query time. Drop-in stores exist for LangChain, LlamaIndex, Haystack and Agno. It is a library rather than a server, and it scans every vector instead of using a graph index.
- Stars
- 17K
- Language
- Rust
- License
- MIT
- Momentum
- Established
DigUp is a free macOS menu bar app that searches the files on your Mac by what is in them, using Google DeepMind's EmbeddingGemma 2 model running locally. You pick the folders it may read. DigUp then indexes pictures, screenshots, PDFs, documents, audio recordings and videos, and you find things by typing a plain description. A hit in a video or recording opens at the matching moment, and a PDF hit opens at the page. Exact strings such as invoice numbers and error codes are matched too, including text that Apple's Vision framework reads from screenshots. Code search is a separate index you switch on yourself: queries start with code: and results open at the line in editors such as Zed, VS Code, Cursor and Xcode. The README says nothing you index leaves the Mac. The app goes online to download the model once (865 MB) and for a daily update check that can be turned off, and it has no account and no telemetry. Internally it runs EmbeddingGemma 2 on llama.cpp with Metal and keeps vectors plus a keyword index in one SQLite file. It needs an Apple Silicon Mac on macOS 14 or later. The latest release is DigUp 0.6.0.
- Stars
- 1.5K
- Language
- Swift
- License
- MIT
- Momentum
- Emerging
oMLX is a local LLM inference server for Macs with Apple Silicon, built on MLX and managed from a menu bar app. It exposes OpenAI-compatible and Anthropic-compatible endpoints, so existing clients can connect to http://localhost:8000/v1 without code changes. Its main idea is a two-tier KV cache. Frequently used cache blocks stay in RAM, and when that hot tier fills up, blocks are offloaded to the SSD in safetensors format so they can be reused instead of recomputed. The author pitches this at long coding sessions with tools such as Claude Code. Concurrent requests go through continuous batching using mlx-lm's BatchGenerator. The server loads text LLMs, vision-language models, OCR models, embeddings and rerankers. Several models can stay loaded at once, with least-recently-used eviction, pinning and a per-model idle timeout. A web admin dashboard at /admin covers monitoring, model management, chat and benchmarks. Source builds also include an experimental mode that splits one model across several Macs. You can install it as a .dmg from GitHub Releases, through Homebrew, or from source, and it is licensed under Apache-2.0.
- Stars
- 23K
- Language
- Python
- License
- Apache-2.0
- Momentum
- Established
Related AI tools
KernelAI
Assistants & chat
An iOS and Android app that runs 53 open-weight models offline on your phone, with no account or subscription.
FreeFree tierVerifiedAgentDock
Coding & dev tools
A Mac app that shows Claude Code, Codex and other coding agents as status dots and lets you answer them from one Inbox.
PaidFree tierVerifiedClaude Code
Coding & dev tools
Anthropic's coding agent that reads a codebase, edits files and runs commands from the terminal, IDE, desktop app or web.
PaidAPIVerifiedRefs
Image & video
Video reference library that turns launch films and explainers into blueprints and prompt recipes Claude and other agents can build from.
FreemiumFree tierVerifiedngram
Image & video
AI video generator that turns a prompt or screen recording into onboarding, training, demo and explainer videos for teams.
FreemiumFree tierAPIVerifiedLune
Writing & research
MCP server that grounds AI agents in full-text computer science papers from top-tier conferences, for students and researchers.
FreemiumFree tierVerified
Related agent skills
Claude Academy guide
anthropics/skills
Recommends a matching course, tutorial or use case from Claude Academy, Anthropic's learning hub, when a user asks how to use Claude or a Claude product.
No scriptsNetwork accessNo hooksNo secretsClaude Code, Claude.ai, Claude API★ 181Kchecked
Discernment nudge
anthropics/skills
Appends two or three follow-up questions after a substantive answer or draft, helping the user check facts, probe reasoning and spot missing context.
No scriptsNo networkNo hooksNo secretsClaude Code, Claude.ai, Claude API★ 181Kchecked
Slack GIF creator
anthropics/skills
Knowledge and Python utilities for making animated GIFs optimized for Slack, covering size constraints, validation tools and animation concepts.
Runs scriptsNo networkNo hooksNo secretsClaude Code, Claude.ai, Claude API★ 181Kchecked