跪拜 Guibai
← All articles
GitHub · Open Source

40 Fresh Open-Source Tools: From 4GB LLMs to AI-Generated PPTs

By HelloGitHub ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The issue captures two converging trends: LLM inference is becoming viable on modest hardware through clever loading strategies rather than brute-force quantization, and AI agent tooling is shifting from raw chat toward structured, reviewable workflows — editable PPTs, annotated diffs, and token-efficient book ingestion — that fit into how teams actually ship software.

Summary

The latest HelloGitHub digest surfaces tools across C, C#, C++, Go, Java, JavaScript, Kotlin, Python, Rust, Swift, and a dedicated AI section. Standouts include airllm, which uses layer-wise loading to squeeze a 70B LLM onto a consumer GPU without quantization or pruning, and oh-my-pi, a terminal coding agent that integrates LSP for refactoring-aware edits and can invoke debuggers like lldb and dlv directly. On the infrastructure side, casos packages Kubernetes into a single binary with a web UI and app store, while dbmate keeps database schemas in sync across teams using plain SQL migrations.

A cluster of AI agent skills shows the ecosystem maturing beyond chat. book-to-skill converts technical books into token-efficient agent skills with auto-generated glossaries and cheat sheets, cutting context cost by 20–50x. dashi-ppt-skill lets an agent produce editable web-based PPTs from a document, and plannotator opens a browser review interface for annotating an agent’s plans and diffs, then feeds the feedback back for execution. Microsoft’s flint-chart introduces a visual intermediate language that lets agents generate accurate charts from concise, human-editable code and exposes an MCP server for direct invocation.

Desktop and productivity tools fill out the list: DeskBox brings native-looking file grids and widgets to Windows, fooyin is a C++ music player with a fully customizable canvas layout, Crisp manages external monitors on macOS for free, and WindInput delivers a Rust-built Chinese input method with Wubi and Pinyin support. A Raspberry Pi digital typewriter and a 3D prehistoric animal museum for kids round out the more playful entries.

Takeaways
airllm runs a 70B model on 4GB VRAM via layer-wise loading, no quantization or distillation required, and supports Llama 3.x, Qwen3, and DeepSeek.
oh-my-pi is a terminal AI coding agent that integrates LSP so renames update all references, and can call lldb, dlv, or debugpy to diagnose issues.
book-to-skill converts PDFs, EPUBs, and DOCX into agent skills with chapter-split files, glossaries, and cheat sheets, reducing token consumption 20–50x versus dumping the full text into context.
plannotator opens a browser review UI for AI-generated plans and diffs; annotations and PR comments feed back to the agent for re-execution, compatible with Claude Code, Codex, and Copilot CLI.
flint-chart is a Microsoft-built visual intermediate language that lets AI agents produce accurate charts from concise code and exposes an MCP service for direct agent calls.
dashi-ppt-skill generates editable web-based PPTs from a document, with in-browser text editing, image swapping, and one-click export to .pptx.
casos bundles Kubernetes API Server, Controller Manager, and Scheduler into a single binary with a web admin UI and app store — no cluster setup or YAML required.
dbmate is a Go-based database migration tool using plain SQL, framework-agnostic, supporting MySQL, PostgreSQL, SQLite, and ClickHouse.
pipecat orchestrates speech recognition, LLM, and TTS into composable pipelines over WebSocket/WebRTC for sub-200ms voice agents, with task handoff and cross-process deployment.
gpu.js transpiles JavaScript functions into GPU shader code for parallel execution across WebGL, WebGPU, and Node.js backends.
tuicr is a terminal code review tool with Vim bindings that lets you browse diffs, add comments, and push the review to GitHub, GitLab, or Bitbucket in one step.
WindInput is a Rust-built Chinese input method supporting Wubi 86, Wubi-Pinyin mixed, full Pinyin, and double Pinyin with low memory footprint.
ServerKit is a Flask+React server panel using 180MB RAM, with 100+ one-click app templates, multi-machine monitoring, and auto HTTPS renewal.
ZenConverter performs local-only audio, video, image, PDF, font, and subtitle conversion on Android — no file uploads to a server.
i-have-adhd is an AI output skill enforcing 10 rules from The Adult ADHD Tool Kit: action-first, numbered steps, max 5-item lists, no pleasantries or recaps.
Conclusions

Layer-wise loading in airllm achieves what quantization and distillation couldn't for consumer GPUs — running a full 70B model without degrading weights — and the approach is model-agnostic enough to cover Llama, Qwen, and DeepSeek families.

The plannotator review loop (agent outputs → browser annotation → feedback re-injected) closes a gap that plain chat interfaces leave wide open: AI-generated code needs the same structured review workflow as human-written code.

book-to-skill's 20–50x token reduction isn't just an efficiency win; it makes domain-specific technical books economically viable as agent context, which changes the calculus for teams sitting on proprietary documentation.

oh-my-pi's LSP integration means the agent doesn't just edit files — it understands symbol references well enough to propagate renames, a capability most terminal coding agents lack and one that reduces the cleanup work after an AI session.

flint-chart treating chart generation as a language rather than a library call suggests a broader pattern: intermediate representations that let agents reason about visual output declaratively may prove more reliable than imperative plotting APIs.

casos collapsing the Kubernetes control plane into a single binary with a web UI lowers the floor for running a PaaS-like environment to the point where a solo developer or small lab can operate one without a cluster admin.

The crop of agent skills (PPT generation, HTML prototyping, ADHD-style output rules, hand-drawn illustrations) signals that the agent ecosystem is building its own 'app store' layer — composable, single-purpose skills rather than monolithic assistants.

Concepts & terms
Layer-wise loading
A technique that loads only the currently active layers of a neural network into GPU memory, offloading the rest to CPU RAM or disk, so a large model can run on a GPU with far less VRAM than the full model requires.
LSP (Language Server Protocol)
A protocol that decouples language smarts (auto-complete, go-to-definition, rename refactoring) from the editor, so any LSP-compliant tool can query a language server for symbol-level code intelligence.
MCP (Model Context Protocol)
An open protocol, pushed by Anthropic, that standardizes how AI agents discover and call external tools, data sources, and services — analogous to a USB-C for agent-tool connections.
DevContainer standard
A specification for defining reproducible, containerized development environments via a devcontainer.json file, supported by VS Code, GitHub Codespaces, and tools like devpod.
WebRTC
A set of browser APIs and protocols enabling real-time peer-to-peer audio, video, and data streaming without plugins, used by pipecat for sub-200ms voice agent interaction.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗