Recently updated open-source projects

Agentless replaces a continuously interactive agent with a three-stage process: fault localization, candidate repair, and patch validation. It narrows an issue from files to classes or functions and then exact edit locations, samples several diff patches, and reranks them using regression and reproduction tests. The project mainly targets SWE-bench-style repair experiments and requires a Python 3.11 environment plus an OpenAI API key.

paper-review.skill guides Claude through a four-stage review of an ML or AI paper: map recent work, test the claimed novelty, write the review, and verify the remaining references. It pushes weaknesses to cite the paper's own evidence and keeps genuine author questions separate from revision requests. It is useful for conference review, pre-submission red teaming, and rebuttal preparation, but is less suited to pure theory, non-English papers, or very narrow fields.

b3sty-skill gives Claude Code and Codex a reusable rule set for RedM and FiveM Lua development. It covers resource structure, server authority, networking, NUI, runtime behavior, databases, debugging, and native calls, with separate guidance for each game. The repository also keeps shared and game-specific working memory, making it suited to implementing or reviewing FXServer resources.

Open Codex CLI forks the original Codex CLI while adding support for OpenAI, Gemini, OpenRouter, Ollama, and compatible endpoints. It uses the Chat Completions API so users can choose a broader range of models, then runs through a familiar terminal workflow after provider credentials are configured. The project is experimental and may change incompatibly.

ClaudeCodeTidy audits project instruction files such as CLAUDE.md, rules, skills, and memory notes, then relocates misplaced content, compresses verbose guidance, and removes only confirmed-dead material. Its companion skill learns from real tidy runs using evidence rather than guesses, while every run records a durable log. The plugin is scoped to instruction hygiene, not application code.

Give the skill a URL and it observes the site in a real browser, then converts its motion tracks, scroll model, assets, and design parameters into a structured specification, an implementation-oriented design manual, and an interactive timeline diagram. It is useful before rebuilding or studying a motion-heavy site, but it requires claude-in-chrome and network access rather than a headless browser.

Superpowers Lab hosts experimental Claude Code skills for tasks that are awkward to handle manually. Current examples include finding semantically duplicated functions, invoking MCP tools on demand without permanent configuration, and controlling interactive programs through tmux; the skills are still being refined and may change.

ghostty-macos-config packages a repeatable macOS Ghostty setup with a Synthwave theme, Maple Mono NF CN, translucent window styling, and terminal shortcuts. It can be applied by Claude or manually, validates the configuration, and writes only Ghostty's own config instead of modifying .zshrc or affecting Terminal.app.

Authme is a cross-platform authenticator for managing time-based two-factor codes on the desktop or web. Add accounts from TOTP QR codes or import existing data from Google Authenticator, 2FAS, or Aegis, then protect the stored codes with password-based AES-256 encryption. It works offline apart from update checks and includes export and backup options for users who want to retain control of their authenticator data.

skill-check validates SKILL.md structure, frontmatter, description quality, body limits, and local references, then reports clear diagnostics and a quality score. It can apply deterministic safe fixes, scan local paths or GitHub repositories, and export JSON, SARIF, HTML, or GitHub annotations. Security scanning is enabled by default in the CLI and may install scanner dependencies unless automatic installs are explicitly disabled.

Claw Code Agent reimplements a Claude Code-style coding workflow in Python for OpenAI-compatible model servers. It provides a CLI, multi-turn chat, a local web interface, file and shell tools, sessions, permissions, plugins, and nested delegation while using only the Python standard library. Development is ongoing, and full MCP, hook, background-team, and platform parity is not yet complete.

SkillCompass is a local-first evaluator and manager for people who maintain agent skills. It scores a skill across six dimensions, identifies the weakest area to improve, and tracks versions so later evaluations show whether a change helped. Usage hooks feed an inbox that surfaces unused skills, stale evaluations, declining adoption, and updates, while SKILL.md edits trigger structural and security-pattern checks. SHA-256 snapshots make regressions recoverable. It is not a general code reviewer, runtime debugger, or tool for creating skills from scratch.

ax ingests local agent sessions, tool calls, plans, skills, Git history, and related signals into a typed graph. It lets you recall past work, inspect skill and tool usage, estimate costs, surface friction, and review proposed changes to project instructions. A session-end retro loop feeds accepted lessons into later work. Ingest is local by default; publishing profiles or sessions is an explicit opt-in.

ga-cli exposes Google Analytics 4 through task-first commands that return flattened JSON and stable exit codes. Its read-only surface covers reports, realtime activity, pages, sources, events, trends, property metadata, compatibility checks, quota checks, and credential diagnostics without interactive prompts.

Upload a LinkedIn PDF, resume, website, or pasted text, answer seven guided questions, and receive a scored section audit, copy-ready rewrites, target keywords, and a short action list. Four modes reweight the advice for branding, job search, creator growth, or consulting. It works only from supplied material and does not edit the live account or promise hiring results.

Paste an existing brief or set of hooks to receive a verdict, problem reframe, strategic score, diagnosis of weak insights, and the hard questions still unanswered. Starting from a blank page, the skill interviews you toward a stronger brief instead of filling a generic template. Its social layer also builds funnel-aware hook matrices with visual, text, and sound delivery for short-form and performance creative.

This skill builds whiteboard explainers in HyperFrames with isolated artwork, serpentine mask reveals, and a photographed hand holding the marker. Whisper word timestamps drive drawings, labels, camera moves, and sound cues, while coverage checks, timed frame audits, and narration verification catch visible failures. It also requires Python, Whisper, Pillow, an image model, and a TTS voice.

Give the skill one focused concept—such as an agent, RAG, or chain-of-thought—and it produces a self-contained HTML lesson with sourced definitions, plain-language explanation, a continuing product example, practice questions, and an interactive demo. It is designed mainly for AI product-manager training and technical education, using a fixed four-step creation process so separate lessons keep a consistent structure.