Weekly digest
This week on HuiaHub: local models, MCP guides and 18 new videos
Local LLMs with Ollama and vLLM, MCP servers in Python, Claude Code hooks, 18 talks and new agent skills: everything HuiaHub published so far.
Most of the new guides are about running models on your own hardware, from Ollama and llama.cpp to vLLM, Whisper and ComfyUI. A second group covers Claude Code itself: connecting MCP servers, writing hooks and authoring a SKILL.md.
New agent skills
- Claude API reference skill is new, with an update later in the week. It gives a coding agent per-language reference files and stops before adding Anthropic SDK calls to code built for another provider.
- Canvas design has the agent write a visual philosophy, then make original artwork as PNG or PDF.
- Document co-authoring takes specs and proposals through three stages: gather context, draft section by section, then check how the result reads.
- Word document tools creates, reads and edits .docx files, with scripts for tracked changes and comments.
Open-source projects
- Agent Skills by Addy Osmani is new and was refreshed later in the week: 25 Markdown workflows and nine slash commands that put a spec, a plan, tests and a review in front of each change.
- DeepSeek's open-source agent runtime turns the model adapter, tools, session log and agent loop into plugins you can swap.
- ECC packages workflows, skills, rules, hooks and security checks for coding agents, with Claude Code best supported.
- Graphify turns a codebase into a queryable knowledge graph locally, without an LLM or API key.
- Understand Anything also maps a repository as a graph, adding LLM-written summaries and guided tours.
- Orca is a desktop app for running several coding agents side by side, each in its own Git worktree.
AI tools
Offrun is a macOS app that runs Claude Code, Codex, AGY and Grok Build through their local CLIs, one Git worktree per agent, and lets a second agent review an uncommitted diff.
Videos
Eight talks come from the AI Engineer World's Fair 2026 in San Francisco:
- Where knowledge should live: prompt, memory or weights. Anant Srivastava of Oracle says to fine-tune reflexes, never facts.
- Serving open models fast. Two Nebius speakers cover cache-aware routing, speculative decoding and quantization.
- Gemini Interactions API and Managed Agents. Ivan Leo of Google DeepMind shows agent state kept on the server.
- A wearable agent on a Raspberry Pi. Jeremy Adams of Neo4j builds it on stage, with graph memory.
- Monitoring LLM apps past the golden signals. Marina Petzel of Datadog adds cost, safety and quality metrics.
- MicroVM sandboxes for coding agents. Rowan Christmas of Docker had his own agent find his bank accounts in five prompts, then reran the attack in a sandbox.
- Why accurate browser agents still fail. Derek Meegan of Browserbase starts from the math: 99% per step gives about 36% over 100 steps.
- Coding agents and the knowledge cutoff. Jakub Hojsan of Exa argues a code review agent needs explicit rules for when to search.
From the Hugging Face channel:
- Structured output in Transformers.js 4.3: a new package holds browser models to a JSON Schema or regex.
- Flow matching for beginners explains how image models turn noise into pictures, with little math.
- Local models in the Pi agent with llama.cpp sets up Qwen3 8B as a coding agent that keeps code on your machine.
- How Mem0 builds agent memory covers its stores, ingestion and hybrid retrieval.
- Prompt caching for agents shows why costs compound over turns and which mistakes break the cache.
Sam Witteveen posted three:
- Image decision models for RPA checks forms and screenshots with two open models and confidence scores.
- Gemini 3.8 Flash TTS covers voice design, voice cloning and Colab demos.
- Nemotron 3 Diarization looks at NVIDIA's model for telling who spoke when, in batch or streaming.
Two more: 3Blue1Brown on cross-entropy explains where the LLM training loss comes from, and Weaviate on agentic RAG traces most failing pipelines to the context they get.
Guides
Running models yourself:
- Run a local LLM with Ollama, call its APIs and size models to your RAM.
- llama.cpp vs Ollama vs LM Studio compares APIs, hardware support and licences, with a decision table.
- Serving an open model with vLLM points the OpenAI SDK at your own GPU.
- Local transcription with Whisper matches openai-whisper, whisper.cpp or faster-whisper to your hardware.
- Open-source text to speech weighs Piper, Kokoro, Coqui TTS and Chatterbox.
- Local image generation with ComfyUI covers vetting custom nodes before installing.
- Object detection with Ultralytics YOLO goes from prediction to training and ONNX export.
Claude Code, MCP and APIs:
- Build an MCP server in Python with the official SDK, then test it in the MCP Inspector.
- Connecting MCP servers to Claude Code explains scopes, sharing through
.mcp.jsonand vetting each server. - Writing a SKILL.md covers frontmatter, descriptions that trigger, scripts and testing.
- Claude Code hooks can format files after edits, run tests and block risky commands.
- Structured outputs with OpenAI and Claude sets out what each API guarantees for JSON.
RAG and evaluation:
- RAG with Postgres and pgvector builds a chunks table, an index and a filtered retrieval query.
- Choosing a vector database compares pgvector, Qdrant, Chroma and Milvus.
- Testing prompts with promptfoo ends with a CI step that fails on regressions.
- Tracing with Langfuse covers traces, token costs and masking personal data.
Sources
- 1.addyosmani/agent-skillsRepositoryPrimarygithub.com
- 2.anthropics/skillsRepositoryPrimarygithub.com
- 3.deepseek-ai/deepseek-harnessRepositoryPrimarygithub.com
- 4.Graphify-Labs/graphifyRepositoryPrimarygithub.com
- 5.stablyai/orcaRepositoryPrimarygithub.com
- 6.Stop Fine-Tuning to Fix Retrieval Problems — Anant SrivastavaWebyoutube.com
- 7.Why 99% Accurate Browser Agents Still Fail — Derek Meegan, BrowserbaseWebyoutube.com
- 8.YOLO Mode, Safely: MicroVM Sandboxes for Any Agent — Rowan Christmas, DockerWebyoutube.com
- 9.Connect Claude Code to tools via MCP - Claude Code DocsOfficial docsPrimarycode.claude.com
- 10.Hooks reference - Claude Code DocsOfficial docsPrimarycode.claude.com
- 11.GitHub - modelcontextprotocol/python-sdk: The official Python SDK for Model Context Protocol servers and clients · GitHubRepositoryPrimarygithub.com
- 12.GitHub - ollama/ollama: Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models. · GitHubRepositoryPrimarygithub.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
- 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.
4 min read
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
Related open-source projects
Understand Anything turns a codebase or supported knowledge base into an interactive knowledge graph for exploring files, functions, classes, dependencies, and relationships. Its code analysis combines Tree-sitter structural parsing with LLM-generated summaries, tags, architectural layers, tours, and business-domain information. The dashboard lets you search the graph, inspect nodes, follow relationships, and view code organized by architectural layer. The main /understand workflow creates .ua/knowledge-graph.json, with older projects continuing to use .understand-anything/. Later runs are incremental by default and re-analyze changed files. Separate commands cover change-impact analysis, explanations of specific code, onboarding guides, domain graphs, and Karpathy-pattern wiki analysis. The project supports Claude Code and a range of other coding agents and IDE integrations. Generated graph data can be committed to a repository and opened by teammates with the standalone viewer. That viewer needs Node.js 18 or newer and reads the graph from local disk without LLM calls. Building or running the analysis code from source requires Node.js 22 or newer and pnpm 10 or newer.
- Stars
- 86K
- Language
- TypeScript
- License
- MIT
- Momentum
- +1.1% 7dEstablished
Orca is a desktop development environment for running multiple AI coding agents side by side, with each task isolated in its own Git worktree. It can launch agents such as Claude Code, Codex, Cursor CLI and OpenCode, while keeping terminals, files, diffs and browser sessions tied to each worktree. The project is aimed at developers who already use coding-agent command-line tools and want one place to manage several sessions. Orca runs on macOS, Windows and Linux. Worktrees can run locally, on SSH machines or on a remote Orca server that you control. Its built-in browser includes a Design Mode that can send an element's HTML, CSS and screenshot to an agent. The CLI can create worktrees, control terminals, open files and diffs, and automate Orca's browser. A mobile companion for iOS and Android can monitor sessions, send replies, inspect files and source control, and receive completion notifications. Orca does not provide an AI model or hosted compute. You bring the coding-agent accounts or subscriptions you already use, and remote execution uses machines or cloud accounts under your control.
- Stars
- 91K
- Language
- TypeScript
- License
- MIT
- Momentum
- +7.8% 7dEstablished
DeepSeek Harness is an open-source agent runtime from DeepSeek AI that runs coding agents through a Web UI, headless jobs, SDK connections or ACP automation clients. Its Cordis-based design treats the model adapter, tools, session log and agent loop as plugins, so these parts can be replaced or extended through configuration. The Web UI works with a selected project directory. An agent can read and edit files, execute commands, delegate work and keep a plan, subject to the configured approval policy. The default setup supports DeepSeek models, while the model settings also support shipped third-party providers and custom APIs using OpenAI Chat Completions, OpenAI Responses or Anthropic Messages protocols. Developers can add their own Harness plugins, including plugins that register tools or services. The project is currently an experimental developer preview. Its maintainers warn that compatibility-breaking changes will occur and that it has not received a security audit. Because agents can execute generated commands and access permitted files, processes, credentials and networks, the safety documentation recommends restricted privileges and isolated environments.
- Stars
- 249K
- Language
- TypeScript
- License
- MIT
- Momentum
- +2.3% 7dEmerging
Graphify turns a software project and related material into a queryable knowledge graph. It parses code locally with tree-sitter, connects structures such as calls, imports and inheritance, and writes the result to graph.json. It can also produce an interactive graph.html viewer and a GRAPHREPORT.md report. Queries can search the graph, explain a concept or trace a path between concepts. Code analysis does not require an LLM or API key. Graphify distinguishes relationships extracted directly from source from relationships resolved by Graphify. It does not use embeddings or a vector store for its graph. Documents, PDFs and images can join the same graph, but their semantic processing uses an AI assistant or configured model backend rather than the code-only parser. The project also supplies a skill that can be registered with supported coding assistants including Claude Code, Cursor, Codex and Gemini CLI.
- Stars
- 126K
- Language
- Python
- License
- Apache-2.0
- Momentum
- +2.0% 7dEstablished
Agent Skills is a repository of 25 Markdown workflows for AI coding agents, covering software work from specification and planning through testing, review, performance, security, deployment and documentation. Each skill is a SKILL.md process with triggers, steps, red flags, verification criteria and common rationalizations to counter. The pack can be installed with the skills CLI or integrated with tools including Claude Code, Cursor, Gemini CLI, OpenCode, GitHub Copilot and Codex. Nine slash commands map development tasks to skills. /build auto can generate a plan and implement its tasks after one approval, while stopping on failures or risky steps. This suits teams that want an agent to follow a defined engineering process instead of relying on isolated prompts. The repository also contains four review personas, shared checklists, evaluation cases and host-specific adapters. One installation limitation matters: installing a single skill through npx does not copy the repository-level references/ directory, so links to shared checklists are unavailable unless those files are supplied another way.
- Stars
- 106K
- Language
- JavaScript
- License
- MIT
- Momentum
- +4.9% 7dEstablished
Related AI tools
Claude 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.
PaidAPIVerifiedOffrun
Coding & dev tools
A macOS workspace for running and reviewing multiple coding agents side by side.
FreeFree tierVerifiedLM Studio
Models & inference
Desktop apps for running open LLMs locally, with a local API server and Bionic, an agent for documents, code and automation.
FreemiumFree tierAPIVerifiedOllama
Models & inference
An open-source tool and API for running open models on your own machine or on Ollama's cloud.
Open sourceFree tierAPIVerifiedLangfuse
Evaluation & observability
Open-source tracing, evaluation and prompt management for LLM apps and agents, as a cloud service or self-hosted.
Open sourceFree tierAPIVerified
Related agent skills
Canvas design
anthropics/skills
Creates original static visual art as PNG or PDF files from a design philosophy, for posters, artwork, designs and similar pieces.
No scriptsNo networkNo hooksNo secretsClaude Code, Claude★ 180Kchecked
Claude API
anthropics/skills
Reference for Claude API and Anthropic SDK work, including models, pricing, parameters, streaming, tools, MCP, agents, caching, token counting and migrations.
Runs scriptsNetwork accessNo hooksAsks for secretsClaude Code★ 180Kchecked
Document co-authoring
anthropics/skills
Guides users through a structured process for writing documentation, proposals, technical specs, decision documents, and similar substantial content.
No scriptsNetwork accessNo hooksNo secretsClaude, Claude Code★ 180Kchecked
Word document tools
anthropics/skills
Teaches an agent to create, read, edit, and manipulate Word documents and templates, including formatting, images, comments, tracked changes, and content extraction.
Runs scriptsNetwork accessNo hooksNo secretsClaude, Claude Code★ 180Kchecked