DEV Community

Cpa Mash
Cpa Mash

Posted on Originally published at xbuff.online

7 GitHub Repos for AI Agents (That Are Actually Worth Bookmarking)

AI agents are only as good as what you give them: a browser to work in, memory to remember with, skills to follow, and a way to organise it all. These 7 open-source repos each solve one of those pieces. None of them is a magic "do everything" agent — each one does one job, and that is why they are useful.

1. Browser Use — give your agent a browser

Repo: https://github.com/browser-use/browser-use

Browser Use is an open-source Python library that lets an AI agent drive a real browser: open pages, click, type, and fill in forms from a plain-language task. It works with OpenAI, Anthropic, Google, and local models, and it is MIT-licensed.

Timely note: version 1.0.0 was released on 2026-11-10 — the first stable release, rebuilt around Chrome's DevTools Protocol (CDP) instead of step-by-step DOM selection.

Start with: pip install browser-use

Best for: research, form-filling, and any task where the agent has to actually use a website, not just read it.

2. Agent Memory — stop re-explaining your project

Repo: https://github.com/rohitg00/agentmemory

Agent Memory is a persistent memory server for AI coding agents. It captures what happens in a session — tools used, files touched, decisions made — and injects the relevant context into the next session, so you stop re-explaining the same project to Claude Code, Cursor, Codex, or Gemini CLI. It works over MCP, REST, and hooks, and it is Apache-2.0 licensed.

Start with: npx @agentmemory/agentmemory, then agentmemory connect claude-code (or your agent of choice).

Best for: anyone who uses a coding agent daily and is tired of starting every session from zero.

3. Scientific Agent Skills — research skills for science agents

Repo: https://github.com/K-Dense-AI/scientific-agent-skills

A library of ready-to-use scientific research skills for AI agents — genomics, cheminformatics, medical imaging, study design, scientific communication, and more. The collection is described in an arXiv paper (September 2026) as 163 procedures in 16 areas; the repo itself has kept growing since, now listing well over 150 skills. It follows the open Agent Skills standard and works with Claude Code, Codex, Gemini CLI, Cursor, and other compatible hosts. MIT-licensed.

Start with: npx skills add K-Dense-AI/scientific-agent-skills

Best for: researchers and students who want an agent that follows real scientific procedures instead of guessing.

4. Diagram Design — diagrams that do not look like default flowcharts

Repo: https://github.com/cathrynlavery/diagram-design

Diagram Design is an agent skill by Cathryn Lavery that turns a plain-language request into a diagram as a single self-contained HTML file with inline SVG — architecture, flowcharts, sequence and state diagrams, quadrants, treemaps, and more, in light, dark, and editorial styles. No Mermaid, no build step, no external images. MIT-licensed.

Start with: npx skills add cathrynlavery/diagram-design

Best for: documentation, architecture explainers, and any post where a clean diagram says more than a paragraph.

5. Anthropic Cybersecurity Skills — security playbooks for agents

Repo: https://github.com/mukul975/Anthropic-Cybersecurity-Skills

A library of over 800 structured cybersecurity skills across dozens of security domains, mapped to industry frameworks including MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, MITRE D3FEND, NIST AI RMF, and the MITRE Fight Fraud Framework. It works with Claude Code, GitHub Copilot, Codex CLI, Cursor, Gemini CLI, and other agentskills.io-compatible platforms. Apache-2.0 licensed.

Two honest notes, straight from the repo itself: this is a community project by Mahipal Jangra — it is NOT affiliated with Anthropic — and the skills include offensive and dual-use techniques meant ONLY for authorised penetration testing, security research, defence, and education, against systems you own or have explicit written permission to test.

Start with: npx skills add mukul975/Anthropic-Cybersecurity-Skills

Best for: security learners and analysts who want an agent with real playbooks, used lawfully.

6. Awesome Harness Engineering — the map, not another tool

Repo: https://github.com/walkinglabs/awesome-harness-engineering

This one is different: it is a curated "awesome list", not a tool to install. It maps the whole field of harness engineering — everything outside the model weights that decides whether an agent actually finishes a job: instructions, tools, environment, state, and feedback. It links to courses and learning resources, foundations, context and memory, constraints and guardrails, agent files and workflow design, evals and observability, benchmarks, and runtimes. The same author also publishes a free course at walkinglabs.github.io/learn-harness-engineering.

Best for: anyone who has watched a strong model fail on a simple task and wants to understand why — the usual answer is the harness, not the model.

7. OpenViking — one place for agent memory, knowledge, and skills

Repo: https://github.com/volcengine/OpenViking

OpenViking, from Volcengine (ByteDance), is an open-source context database for AI agents. Instead of separate systems for retrieval, conversation memory, and skill management, it stores memory, resources (docs, repos, pages), and skills as one virtual filesystem under viking:// URIs, which the agent browses with ordinary file operations like ls, tree, and find. Launched January 2026. Core is AGPL-3.0 licensed; CLI crates and examples are Apache-2.0.

Start with: git clone https://github.com/volcengine/OpenViking.git — and read the README before running it, because setup depends on your model provider.

Best for: builders who are tired of stitching three different memory systems together.

How to use this list

Do not install all 7 at once. Pick the one gap that hurts most right now: Agent cannot use the web? Start with Browser Use. Agent forgets everything between sessions? Start with Agent Memory or OpenViking. Agent makes ugly diagrams? Diagram Design. Agent needs real procedures, not guesses? Scientific Agent Skills — or the cybersecurity library, if security is your field. Agent keeps failing on long tasks? Read Awesome Harness Engineering before swapping models. One repo at a time, one real task to test it on, and keep the one that earns its place.

These are all free and open source, and all 7 were live and actively maintained when checked on 2026-11-11. That is the point: the agent ecosystem is moving fast — Browser Use shipped 1.0.0 yesterday — so bookmark the repos, not just this article, and check the README before you install anything.


Originally published at xbuff.online. More plain-English AI guides and tools at xbuff.online — prompt packs and 100 n8n automation workflows in the shop.

Top comments (0)