# AI Agent Tools — Research Notes

A quick reference on ten open-source AI agent projects. All repos were verified as real at time of writing.

---

## 1. Goose
**Repo:** [`block/goose`](https://github.com/block/goose)
Open-source, extensible AI coding agent originally built by Block (Square/Cash App), now an anchor project of the Linux Foundation's Agentic AI Foundation. Runs as both a CLI and desktop app, works with any LLM (including fully local models via Ollama), and goes beyond suggestions — it can install, execute, edit, and test code. Apache-2.0.

## 2. Page Agent
**Repo:** [`alibaba/page-agent`](https://github.com/alibaba/page-agent)
A JavaScript in-page GUI agent from Alibaba. You drop a script into a web page and an AI agent controls the interface with natural language — clicking, typing, selecting, and scrolling. Runs entirely client-side in the browser: no server, no Python, no headless browser required. MIT.

## 3. Agent Orchestrator
**Repo:** [`ComposioHQ/agent-orchestrator`](https://github.com/ComposioHQ/agent-orchestrator)
An orchestrator for running multiple coding agents in parallel. It plans tasks, spawns agents in isolated workspaces, and autonomously handles CI fixes, merge conflicts, and code reviews — each producing pull requests a human reviews and merges. Focused on software development workflows where many agents work on different issues simultaneously.

## 4. Hermes Agent
**Repo:** [`NousResearch/hermes-agent`](https://github.com/NousResearch/hermes-agent)
A self-improving AI agent from Nous Research, built around a learning loop: it creates skills from experience, refines them during use, persists knowledge, searches its own past conversations, and builds a model of the user across sessions. Works with any OpenAI-compatible API (OpenRouter, OpenAI, Anthropic, Google, local models, etc.). Requires a Unix-like environment (WSL2 on Windows).

## 5. Claude Code Plugins
**Repo:** [`anthropics/claude-code`](https://github.com/anthropics/claude-code) (plugins directory)
Plugins are extensions for Anthropic's Claude Code that add custom slash commands, specialized sub-agents, hooks, and MCP servers. They can be shared across projects and teams for consistent tooling and workflows. Anthropic also maintains an official plugin directory ([`claude-plugins-official`](https://github.com/anthropics/claude-plugins-official)) and a community marketplace.

## 6. Copilot SDK
**Repo:** [`github/copilot-sdk`](https://github.com/github/copilot-sdk)
A multi-platform SDK from GitHub that exposes the same agent runtime behind Copilot CLI, so you can invoke it programmatically. You define agent behavior and Copilot handles planning, tool invocation, and file edits. Ships language SDKs (TypeScript/Node.js, .NET, Java, Python, Go, Rust) that talk to the Copilot CLI over JSON-RPC.

## 7. Symphony
**Repo:** [`openai/symphony`](https://github.com/openai/symphony)
An open-source spec/engineering-preview from OpenAI for Codex orchestration. It turns a project-management board like Linear into a control plane for coding agents: it continuously reads open work items, creates an isolated workspace per issue, runs an agent session, and returns proof of work (CI status, PR feedback, walkthroughs). Explicitly labeled a preview for trusted environments, not a maintained product.

## 8. HiClaw
**Repo:** [`alibaba/hiclaw`](https://github.com/alibaba/hiclaw)
An open-source collaborative multi-agent OS. Multiple agents coordinate through Matrix chat rooms with full human visibility and real-time intervention — a Manager Agent coordinates Worker Agents to complete complex tasks. Emphasizes transparent, auditable, human-in-the-loop task coordination. (Also mirrored under the AgentScope/Higress orgs.)

## 9. NotebookLM-py
**Repo:** [`teng-lin/notebooklm-py`](https://github.com/teng-lin/notebooklm-py)
An unofficial Python API, CLI, and agentic skill for Google NotebookLM, built on undocumented RPC APIs. Gives programmatic access to features — including some the web UI doesn't expose: create notebooks, add sources (URLs, YouTube, PDFs, audio, video, images), chat with content, and generate artifacts (podcasts, videos, quizzes, infographics). Uses browser-based OAuth (no stored passwords) and integrates with agents like Claude Code and Codex.

## 10. OpenSandbox
**Repo:** [`alibaba/OpenSandbox`](https://github.com/alibaba/OpenSandbox)
A general-purpose sandbox platform for AI applications. Provides multi-language SDKs, unified sandbox APIs, and Docker/Kubernetes runtimes for isolated code execution. Targets coding agents, GUI agents, agent evaluation, AI code execution, and RL training. Includes a FastAPI control plane to create, run, monitor, and dispose of containerized sandboxes.

---

*Descriptions paraphrased from each project's README/homepage for licensing compliance. Timestamps in the original list appear to reference a video walkthrough.*