Shape agent capabilities into your workflow.

Agent Lightning connects existing agents to optimization workflows with minimal code changes by collecting prompts, tool calls, and rewards as structured traces. It can selectively optimize one or more agents built with LangChain, the OpenAI Agent SDK, AutoGen, CrewAI, Microsoft Agent Framework, or plain Python. Supported approaches include reinforcement learning, automatic prompt optimization, and supervised fine-tuning.

Apply a focused set of skills to context degradation, compression, memory design, tool interfaces, multi-agent coordination, and long-horizon task execution. The repository also covers operational concerns such as evaluation, LLM-as-a-judge methods, harness safeguards, and self-improvement loops. Complete examples show how several skills can be combined into larger agent systems across different hosts.

Oh My Pi is a terminal coding agent that works across many model providers and uses optimized reading, search, and hash-anchored editing tools. It brings LSP operations, real debugger sessions, browser access, persistent Python and Bun execution, and isolated subagents into one interface. It suits developers who want complex coding, review, and automation workflows without leaving the command line.

Gas Town manages a workspace where multiple coding agents can work across projects without losing their state when sessions restart. Git-backed hooks, persistent identities, mailboxes, handoffs, and a Beads ledger keep work traceable, while project containers and worker roles organize parallel tasks.

Plandex is a terminal coding agent designed for large projects and real development tasks. It loads only the context needed for each step, generates and reviews changes across many files, and can debug builds, tests, deployments, or scripts. Users can choose between full autonomy and fine-grained control or combine models from different providers; Windows is supported only through WSL.

SurfSense gives competitive-intelligence agents live connectors for social platforms, search, maps, and the open web through REST APIs or MCP. Scheduled and event-triggered agents can combine sources into cited briefs and alerts, then keep findings in a searchable knowledge base. The platform, including connectors, automations, and the MCP server, can be self-hosted with Docker, but the project states that it is still under active development and not yet production-ready.

The Self-hosted AI Starter Kit combines n8n, Ollama, Qdrant, and PostgreSQL in a preconfigured Docker Compose environment. It provides a quick foundation for local agents, private document summarization, Slack bots, and other low-code AI workflows. On Apple Silicon, Docker cannot use the Mac GPU directly, so users must run the stack on CPU or host Ollama on macOS and connect it to the containers.

Awesome Codex Skills organizes practical skills across development, productivity, communication, data analysis, and utility categories, often with a short use case and installation command. The catalog mixes skills stored in the repository with links to external projects, so it is useful for discovery but each item’s permissions, dependencies, and quality still need separate review.

This repository collects Agent Skills for Google products and technologies, with a substantial set of Google Cloud workflows for onboarding, architecture, and AI/ML tasks. The installer lets you choose individual skills rather than enabling the entire collection, and the repository is still under active development.

OpenHarness is a Python framework for researchers and developers who want an inspectable, extensible agent runtime through a CLI or terminal UI. It combines tool use, skills, persistent memory, permissions, MCP, and multi-agent coordination. The bundled ohmo agent can work through Feishu, Slack, Telegram, or Discord using an existing Claude Code or Codex subscription.

Skill Seekers uses a Python CLI and MCP server to turn documentation sites, GitHub repositories, PDFs, videos, and other sources into structured knowledge assets. It can compare code with documentation, generate example-rich SKILL.md files, and export for Claude, Gemini, OpenAI, LangChain, and multiple vector-database formats. It also supports caching, resumable jobs, and project-configuration scanning. Basic use requires Python 3.10+ and Git; AI enhancement or upload features may need provider credentials.

Run an LLM agent that can operate a real browser, terminal, files, keyboard and mouse, screen vision, and Android devices through a small set of local tools. GenericAgent records successful task paths as reusable skills, but you must supply a model API key and complete extra local setup for capabilities such as browser automation, vision, or computer control.

NVIDIA SkillSpector examines Git repositories, URLs, archives, directories, or individual files before an agent skill is installed. It checks dozens of patterns spanning prompt injection, exfiltration, privilege escalation, dangerous code, dependency vulnerabilities, MCP least privilege, and tool poisoning, then produces terminal, JSON, Markdown, or SARIF reports with risk scores. Scanned skills are never executed. Optional LLM analysis sends file contents to the configured provider, while static-only mode keeps that content local.

CloudCLI adds a responsive web interface to Claude Code, OpenCode, Cursor CLI, and Codex, letting you resume sessions, chat, use a terminal, edit files, inspect Git changes, and run browser tasks from desktop or mobile. The open-source edition self-hosts through npx on Node.js 22 and discovers existing sessions automatically. Experimental Docker sandboxes offer stronger isolation, while a managed cloud option is also available.

IronClaw runs a personal AI assistant through a CLI, web interface, and messaging channels while keeping its state under your control. It supports multiple model providers, MCP connections, scheduled routines, and extensible tools, with WASM sandboxing and credential isolation for untrusted operations. Building the full system from source requires a Rust and PostgreSQL environment.

Superset coordinates command-line coding agents across isolated Git worktrees, with status monitoring, notifications, a built-in diff viewer, and one-click handoff to an editor or terminal. It works with several named agents and can accept any agent that runs in a terminal. Local development currently requires Bun, Docker, and Caddy, and downloadable Windows and Linux builds are not yet available.

Microsoft Agent Framework targets Python and .NET applications that need more than a single prompt-and-response loop. It supports sequential, concurrent, handoff, and group-collaboration graphs, along with checkpointing, streaming, human-in-the-loop control, OpenTelemetry tracing, and Foundry hosting samples. Its scope suits teams building restartable, observable agent workflows rather than a small one-off script.

BISHENG is a self-hosted platform for creating and operating enterprise LLM applications. It brings workflow orchestration, RAG, agents, model and dataset management, evaluation, fine-tuning, administration, and observability into one browser-based environment. Deployment uses Docker Compose; the documented minimum is four virtual CPU cores and 16 GB of RAM, and the default stack also installs Elasticsearch, Milvus, and OnlyOffice.