Start with the job, then find an open-source answer

Termul treats each project as a persistent desktop workspace instead of a loose set of terminal tabs. You can combine multiple shells, split panes, embedded browser tabs, and code or Markdown editors, then restore sessions, layouts, and snapshots when you return. The application is packaged for Windows, macOS, and Linux and also supports SSH-oriented workflows with credentials stored through the operating system keychain. It is designed for developers who regularly switch among several projects and want their working context kept together.

Canvas MCP gives students and educators more than ninety tools for Canvas LMS, covering assignments, to-do items, course information, grades, teaching workflows, and bulk grading. Five Agent Skills package common student and educator routines. It can be installed as a Claude Desktop extension or a standard MCP server and requires the institution's Canvas API URL plus a personal access token.

Skills Manager brings GitHub skill catalogs and local folders into a three-column macOS app with search, tag filtering, Markdown previews, editing, installation, and removal. It can manage both Claude Code and Codex skill locations and supports multiple catalogs, but requires macOS 15 or later.

Install it in a project to generate agent instructions, reusable skills, a dependency graph, and Git hooks for tools such as Claude Code, Cursor, and Gemini. The setup helps agents answer structural questions and enforces rules around commits, file headers, and complexity; in-agent skills need no separate API key, while standalone CLI features may require one.

Firelink is a cross-platform desktop download manager for segmented transfers, media extraction, persistent queues, scheduling, speed rules, and organized file placement. Its companion extension can hand browser downloads, links, selected text, or detected media to the desktop app for review, and release bundles include the required engines. Builds are available for macOS, Windows, and Linux, but the current macOS package is not notarized and Windows packages are unsigned, so first launch may trigger security warnings.

Inline Studio is a cross-platform desktop app for organizing an AI film on a free-form node canvas, from moodboards and generated shots to trimming and final assembly. Every render is preserved as a non-destructive take, and frames can feed downstream nodes for repeatable iteration. Generation uses either your own reachable ComfyUI instance or hosted fal models with your own key; current installers are unsigned.

This command-line scanner reviews third-party Claude Skills before installation for suspicious network access, sensitive file operations, dangerous commands, dynamic execution, and injection patterns. It can produce HTML, console, or JSON reports and supports custom rules and allowlists, while performing static analysis without executing the skill code. Its findings are a review aid rather than a guarantee that every security risk will be detected.

Codex Usage Tracker reads the JSONL logs already stored on your machine, indexes aggregate call, thread, cache, and supporting evidence into SQLite, and exposes the results through a localhost dashboard, CLI, and MCP tools. You can filter costly or low-cache calls, inspect thread-level drain, and ask Codex to explain usage patterns. Indexed content remains local, while generated reports and support bundles omit it by default. The project is independent and not affiliated with OpenAI.

Taurine is a starter for combining a Tauri desktop shell with a Node.js server compiled into a sidecar binary. It gives web developers a ready React, Vite, Fastify, and tRPC setup that can be launched or built with the usual Tauri commands. The repository is explicitly experimental rather than a production foundation: it relies on a localhost protocol that Tauri discourages for security reasons, may not work on every platform, and only builds for the operating system running the build.

This sample demonstrates how a Windows WPF application uses MSAL.NET to obtain an access token and call Microsoft Graph through the Microsoft identity platform. It is preconfigured for a quick run in Visual Studio, while developers can replace the supplied coordinates with their own app registration. Its token cache uses DPAPI, so the cache file must be removed before moving a built application to another machine.

Pawn Appétit lets players import games from Lichess, Chess.com, or PGN files, analyze positions with Stockfish or other UCI engines, and practice opening repertoires with spaced repetition. Prebuilt releases are available for Windows, macOS, and Linux, and its optional telemetry can be disabled in settings.

This repository packages Agnes free-model access as three independently installable skills for text, images, and video. Local Python helpers cover chat and streaming responses, image generation and transformation, and asynchronous video creation with status polling and downloads. An Agnes API key must be supplied through environment variables, while the repository deliberately excludes real credentials, generated media, and local outputs. Users can install only the capability they need.

This tool controls a signed-in Windows WeChat client through UI Automation to send text, image URLs, local images, and files to named contacts or groups. Agents can use the included skill or one-shot CLI without running a background service, while longer-lived integrations can use the Flask HTTP API and message queue. It does not hook the client or implement the WeChat protocol, but the desktop client must stay signed in and the Windows UI Automation and clipboard environment must remain available; recipient names must match WeChat exactly.

softspark ai-toolkit distributes a broad set of skills, 44 specialist agents, lifecycle hooks, shared rules, personas, and optional plugin packs across many AI coding environments. It supports planning, debugging, review, TDD, commits, pull requests, documentation, orchestration, and parallel workflows, with adapters for Claude, Cursor, Windsurf, Gemini, Copilot, Codex, OpenCode, and others. Executable hooks can block selected destructive operations and enforce quality or session policies, although platforms without hook support receive guidance rather than equivalent machine enforcement.

agent-qa lets teams describe web and mobile tests in natural language, then re-observes the interface and tries another path when an action drifts or fails. It learns from previous runs, exposes a CLI and dashboard, supports MCP and multiple model providers, and can run setup or teardown hooks inside Docker containers. Docker is required for those hooks, and the harness remains an aid to a team's own release checks.

Frigoligo is a cross-platform wallabag client for reading and synchronizing saved articles on Android, iOS, macOS, Linux, and Windows. Its offline-first design keeps reading available without a network connection and synchronizes changes when possible. Deep links and operating-system integrations let external workflows open a particular article, save a new URL, or navigate to app actions. Deep-link support is available on every listed platform except Linux. The app is a client rather than a standalone read-later service, so it requires access to a wallabag server.

VitaBench evaluates agents in real-world service scenarios such as delivery, in-store operations, and OTA, using databases, API tools, and 100 tasks per domain. Configure the models, run the vita command, save simulations under data/simulations, and re-evaluate existing runs when needed. The current release updates datasets, tools, evaluator models, and metrics, so scores should be interpreted alongside the chosen configuration.

M-Courtyard provides a graphical, local workflow for preparing data and fine-tuning language models on Apple Silicon. Users can import TXT, PDF, or DOCX files, clean and segment them, generate instruction datasets, run MLX-based LoRA training, and test the adapter in a built-in chat. Models can be exported for Ollama, mlx-lm.server, or LM Studio. The application requires macOS 14 or later and an M1, M2, M3, or M4 Mac, with local Python and mlx-lm as the core runtime; 16 GB or more memory is recommended for 7B or 8B models.