Recently updated open-source projects

Caliber analyzes a repository and creates tailored context files, rules, skills, and MCP settings for tools such as Claude Code, Cursor, Codex, OpenCode, and GitHub Copilot. After setup, refresh hooks keep those files aligned with code changes. Proposed edits are shown as diffs and backed up, while generation uses an existing agent subscription or a user-provided model key.

llm-wiki-skill has an agent compile raw material into a persistent Markdown knowledge base, handling writing, filing, cross-references, logs, and maintenance over time. Companion Obsidian and browser tools let users attach feedback to selected text and track open audits. The repository labels the skill experimental and expects it to evolve.

Connect OpenConcho to one or more self-hosted Honcho instances to inspect workspaces, peers, sessions, conclusions, and consolidation results from a visual interface. It also supports semantic conclusion search, memory-aware chat, webhooks, and cross-instance fleet views. Desktop builds are available for macOS, Windows, and Linux, with web and container deployments as alternatives; a running Honcho instance is required, and connection details remain in the browser.

aria2-ariang-docker packages Aria2, the AriaNg web interface, File Browser, and Rclone into a container for remote downloading and file handling. Users can browse or play stored files, enable basic authentication and HTTPS, and automatically upload completed downloads to configured cloud storage. It supports ARM systems such as Raspberry Pi, but persistent configuration and session volumes are important if download progress must survive container restarts.

Blueprint packages a focused software-delivery process into agent skills for design documents, specifications, planning, implementation, testing, browser verification, and code review. Teams can invoke one stage or coordinate tickets through review-ready pull requests. Completion is tied to tests and observable proof, and browser-based checks require a compatible automation setup.

visualise teaches an agent to produce self-contained HTML and SVG for clickable diagrams, live charts, comparison views, steppers, and other explainers. Clients that render its visualizer code fence can display the result inline in a sandboxed frame. In unsupported clients, the generated HTML remains usable as a standalone browser file.

agent-skills-eval tests an Agent Skill against the same prompts in two modes: with the skill loaded and without it as a baseline. A judge model grades both outputs against defined assertions, producing JSON and JSONL artifacts plus a static HTML report that shows measurable lift, failures, timing, and tool-call behavior. The project includes a command-line runner and a TypeScript SDK, supports OpenAI-compatible chat backends and custom providers, and can run in CI. Users supply both the target model and the judge model.

TweakPHP provides a focused desktop code editor for modifying and experimenting with PHP code. Builds are distributed for macOS, Windows, and several Linux package formats, including Homebrew, Snap, AppImage, RPM, Deb, and AUR. The project is explicitly labeled as beta, so it is best suited to users comfortable testing an early-stage application.

skill-validator is a CLI for checking Agent Skill packages beyond basic specification compliance. It validates structure and frontmatter, detects broken or orphaned references, flags unnecessary files, reports token usage, and analyzes instruction density and cross-language contamination. Optional model-based evaluation can score dimensions such as clarity, actionability, and novelty. The tool can be installed through Homebrew or Go, used as a pre-commit check, integrated into CI, or imported as a Go library for custom validation workflows.

pear-rec is a cross-platform desktop tool for taking screenshots and recording the screen, audio, video, or animated GIFs. Its screenshot workflow includes cropping, drawing, shapes, color tools, filters, QR recognition, clipboard export, and history, while recordings are handled through WebRTC and WebCodecs-based components. The feature checklist may include work completed in development but not yet shipped in the latest release, so release notes should be checked before relying on a specific function.

Stlite runs Streamlit applications inside the browser through Pyodide and WebAssembly, removing the need for a Python server. Apps can be mounted with a custom HTML element or JavaScript API, including multiple files, package requirements, configuration, and optional IndexedDB persistence. Because execution stays in a browser sandbox, packages with unsupported native extensions and some blocking Streamlit behaviors will not work as they do in a standard Python environment.

Repo Analyzer scans a project, asks the user to choose an analysis depth, and can dispatch parallel subagents across core modules. It emphasizes why the architecture was designed a certain way, compares trade-offs with alternatives, and saves a single Markdown report with Mermaid diagrams. External search MCP servers are optional rather than required.

This bilingual skill teaches engineers how to design the systems surrounding an AI model loop. Six chapters and eleven deep-dive references cover memory, skill loading, tools and permissions, context engineering, multi-agent coordination, and lifecycle design, with tradeoffs and common failure modes. The patterns were distilled from source-level analysis of the Claude Code runtime, but they are presented as runtime-agnostic principles for people building coding agents, plugins, or orchestration systems—not as a source mirror or a beginner user guide.

This collection provides Agent Skills and plugins for Expo and React Native development. Installed through supported Codex, ChatGPT, or Claude plugin flows, the workflows can generate platform-aware app icons, scaffold projects from templates, create local Android and iOS release artifacts, and add a shared cross-platform theming system. The repository is aimed at mobile developers who want coding agents to handle repetitive setup and platform-specific details. Individual plugins may still require their companion CLI, an OpenAI-backed image service, or local Android and Xcode build tooling.

Point a coding agent at this repository to identify prompt-cache fixes that match the harnesses you actually use, apply the cited diffs, and run each skill's verification steps. The bundle covers targeted issues in tools such as Aider, Cline, Roo Code, Continue, and OpenCode, backed by separate audit notes. It is not a universal patch set: some audited tools, including Codex CLI and the tested Claude Desktop Code baseline, are documented as already caching correctly.

The repository currently provides the `codex-dynamic-workflows` skill for supervised AI-agent workflows. It can organize a goal into subagents or simulated work packets, pause at approval gates, integrate results in the parent flow, verify the outcome, and retain reusable workflow artifacts. Users can install the skill in a compatible agent and invoke it when a task benefits from staged coordination and explicit oversight.

Give mimeo a person’s name and it searches public essays, talks, interviews, and other sources, distills recurring principles, and writes a SKILL.md, an AGENTS.md file, or both for a coding agent. It can verify extracted quotes and refine drafts through critique passes. Parallel and OpenRouter API keys are required, with extra dependencies for local podcast transcription.

TipTour is a macOS menu-bar companion that turns spoken requests into computer actions. Hold a hotkey and it can use the current window, hovered element, or a freeform highlighted area to click, type, open apps, scroll, or revise selected text. Gemini Live handles voice and screen understanding, while native accessibility data and its control driver target interface elements. Source builds require macOS 14 or later, Xcode 16 or later, the user's own Gemini API key, and microphone, screen-recording, and accessibility permissions.