GitHub repos for AI agents: 5 picks for week 1
Guide/ 10 min read
superpowers, claude-mem, agent-skills, impeccable and cmux: what each one solves, how to install it, its license, and what it runs on your machine.
- AI
- Claude Code
In this article (5)
These are five GitHub repos for AI agents worth knowing this week: obra/superpowers, thedotmack/claude-mem, addyosmani/agent-skills, pbakaus/impeccable and manaflow-ai/cmux. For each one we checked, on October 30, 2026, what it solves, its license, how active it is, how it installs and what access it asks for on your machine.
We did not install any of them to write this: we read their READMEs, their plugin manifests and the GitHub API data, and the install commands are the ones each project publishes. If you already installed Anthropic's official skills, these are the next step: community projects with more reach, and more risk.
The 5 GitHub repos for AI agents and what each one solves
Four of them are Claude Code plugins: they add skills, commands and, in three cases, hooks that run on their own. The fifth is an application, a macOS terminal. They do not compete with each other: each covers a different job.
Scroll the figure sideways to see all of it
obra/superpowers: a full way of working
obra/superpowers calls itself an agentic skills framework and a software development methodology. When you ask for something, the agent does not jump into code: it asks what you are really trying to do, shows you the design in short chunks, writes a plan of 2–5 minute tasks and executes it with TDD (failing test first) and reviews between tasks. The skills trigger on their own. MIT license; version 7.0.1 shipped on October 21, 2026.
/plugin install superpowers@claude-plugins-official
- It installs from Anthropic's official marketplace, which Claude Code adds by itself; its entry there points at a specific commit of the repository. The README also offers its own marketplace,
obra/superpowers-marketplace. - Use case: multi-step features where you want spec, plan and test discipline. For fixing a typo it is too much process; that is our take, not the project's.
- Permissions: it registers a
SessionStarthook that runs a script with bash when a session starts, is cleared or compacts. Its optional visual companion loads a logo from the project's website with the version number;SUPERPOWERS_DISABLE_TELEMETRYturns that off.
thedotmack/claude-mem: memory across sessions
thedotmack/claude-mem captures what your agent does during a session, compresses it with an AI model and injects the relevant parts into future sessions. It runs on a local worker service (an HTTP API with a web viewer), SQLite, a Chroma vector database and an MCP server with search tools. It needs Node 20 or newer and installs Bun and uv if they are missing. Apache-2.0 license; version 15.7.0 shipped on October 30 itself.
npx claude-mem install
- This is the installer the README recommends. At the end it asks you to sign in to claude-mem's hosted service, with a free trial, or to pick another provider: your own OpenRouter or Gemini key, or your Anthropic plan. To skip the sign-in, it documents the
--providerflag and theCLAUDE_MEM_ONLINE_OPTIN=falsevariable. - Inside Claude Code it also installs with
/plugin marketplace add thedotmack/claude-memand/plugin install claude-mem. Careful:npm install -g claude-meminstalls only the library, with no hooks and no worker. - Use case: long projects where every new session starts by re-explaining the same things.
- Permissions: hooks on 9 events, including
UserPromptSubmit,PreToolUseandPostToolUse; a local MCP server; and what it compresses goes to the AI provider you choose. To keep something out, the README points to<private>tags.
addyosmani/agent-skills: the engineering lifecycle as commands
addyosmani/agent-skills packages 25 engineering skills and 9 commands that follow the development lifecycle: /spec, /plan, /build, /test, /constraints, /review, /webperf, /code-simplify and /ship. Every skill ends by demanding evidence: passing tests, build output, real data. MIT license; version 0.6.14 from October 29, 2026.
/plugin marketplace add addyosmani/agent-skills
/plugin install agent-skills@addy-agent-skills
- The first line adds the repository as a marketplace; the second installs the
agent-skillsplugin from theaddy-agent-skillsmarketplace, the name itsmarketplace.jsondeclares. - Claude Code already has a built-in
/review, so this plugin's version is called by its full name:/agent-skills:review. - Use case: teams that want explicit gates between specifying, building, reviewing and shipping.
- Permissions: the plugin registers no hooks by itself: there is no
hooks/hooks.json. The scripts in itshooks/folder are opt-in and you wire them into yoursettings.jsonby hand. Of the four plugins, it is the one that runs the least on its own.
pbakaus/impeccable: design vocabulary and a detector
pbakaus/impeccable gives your agent words to steer design with: 24 commands such as /impeccable critique, /impeccable audit or /impeccable polish, plus a detector with 62 deterministic rules that flag the usual tells of an interface generated without judgment. The detector uses no model and no API key. Apache-2.0 license; CLI version 4.8.0 from October 26, 2026.
npx impeccable install
- The installer detects your tools, asks whether to install for the project or globally, and offers the design hook. Then you run
/impeccable initonce per project, which writes aPRODUCT.md. - Claude Code also has the marketplace route:
/plugin marketplace add pbakaus/impeccable, then install it from/plugin. - Use case: reviewing and polishing interfaces that already exist. Our own suggestion: pair it with Anthropic's frontend-design skill, one guiding the build and the other reviewing it.
- Permissions: hooks on
SessionStart, onPostToolUseafter everyEditorWrite, and onStop; all three run the project's own binary, which ships with the skill or downloads once into~/.impeccable/bin/. Its live mode runs a script from yourpackage.jsonwith your permissions.
manaflow-ai/cmux: a terminal for many agents
manaflow-ai/cmux is a native macOS terminal, built in Swift on Ghostty's rendering engine, designed for running several agents at once: vertical tabs that show the git branch and PR, notifications when an agent is waiting for you, a built-in browser the agent can drive, and SSH workspaces. It runs on macOS only. Version 0.65.2.2 from October 27, 2026.
brew tap manaflow-ai/cmux
brew install --cask cmux
- The first line adds the project's Homebrew tap; the second installs the app. The README also offers a DMG, and the app updates itself through Sparkle.
- Use case: you have three or four Claude Code sessions open and lose track of which one is waiting on you.
- License: GitHub shows
NOASSERTIONbecause the file combines two: the app and CLI are GPL-3.0-or-later, and the server directories (web/and the workers) use the Business Source License 1.1, which requires a commercial license for production use. - Permissions: it is your terminal, so everything that runs in it has your shell. Its CLI and socket can send keystrokes, read the screen and drive the browser; it can import cookies from your browsers if you use that feature.
Scroll the figure sideways to see all of it
How to evaluate a GitHub repo before you install it
A plugin is not a harmless list of instructions. The Claude Code plugin security page says it plainly: a plugin can execute arbitrary code with your user privileges, and its hooks and MCP servers run outside the sandbox, without going through the permission rules that govern what Claude does. Before installing, check three things.
License and activity
The GitHub CLI gets both in one line, without cloning anything. It works the same in Bash and PowerShell:
gh api repos/obra/superpowers --jq "{license: .license.spdx_id, archived, pushed_at, stars: .stargazers_count}"
{"archived":false,"license":"MIT","pushed_at":"2026-10-29T23:35:34Z","stars":309369}
gh api repos/obra/superpowersasks the GitHub API for the repository's data. You need theghCLI, signed in.--jqfilters the response:.license.spdx_idis the license identifier,archivedsays whether the owner froze it,pushed_atis the date of the last push andstargazers_countis the star count.- The second line is the real response from October 30, 2026: MIT, active, last pushed the day before.
If the license comes back as NOASSERTION, as it does for cmux, GitHub could not classify it: open the LICENSE file and read it. This is what the five returned that day:
| Repo | License | Last push | Latest release | Stars (Oct 30, 2026) |
|---|---|---|---|---|
| obra/superpowers | MIT | Oct 29, 2026 | v7.0.1 | 309k |
| thedotmack/claude-mem | Apache-2.0 | Oct 30, 2026 | v15.7.0 | 106k |
| addyosmani/agent-skills | MIT | Oct 29, 2026 | 0.6.14 | 112k |
| pbakaus/impeccable | Apache-2.0 | Oct 30, 2026 | cli-v4.8.0 | 97k |
| manaflow-ai/cmux | NOASSERTION (GPL-3.0-or-later and BUSL-1.1) | Oct 30, 2026 | v0.65.2.2 | 28.6k |
None of them is archived. Star counts change by the minute: they tell you many people use a project, not whether it suits you.
Permissions: what runs on its own
What matters is what the plugin executes without you asking. For a Claude Code plugin, look at three things in its folder: hooks/hooks.json (the commands that run on each event), .mcp.json (the MCP servers it starts) and bin/ (executables added to the PATH). You can read them without cloning too:
gh api -H "Accept: application/vnd.github.raw" repos/obra/superpowers/contents/hooks/hooks.json
{
"hooks": {
"SessionStart": [
{
"matcher": "startup|clear|compact",
"hooks": [
{
"type": "command",
"command": "\"${CLAUDE_PLUGIN_ROOT}/hooks/run-hook.cmd\" session-start",
"shell": "bash",
"async": false
}
]
}
]
}
}
- The
Accept: application/vnd.github.rawheader makes the API return the file as is, instead of base64-encoded. SessionStartwithstartup|clear|compact: the hook runs when a session starts, is cleared or compacts.commandis what gets executed, with bash. Your next step is to openhooks/run-hook.cmdin the repository and read what it does.
Repeat it with each plugin's path: in claude-mem and impeccable the hooks live in plugin/hooks/hooks.json, because their marketplace points at the plugin folder. After installing, claude plugin details prints the inventory of hooks, MCP servers and skills for the installed copy. Here is a summary of what we found:
Scroll the figure sideways to see all of it
Common mistakes
- Installing by star count. 300k stars say a project is popular, not what it runs on your machine. Read the hooks first.
- Installing superpowers and agent-skills together without deciding which one leads. Both bring their own plan, test and review method, and both trigger skills on their own. Try one; if you add the other, decide which commands you use from each.
- Accepting an installer's sign-in without reading it. claude-mem's offers a hosted service; if your code cannot leave your machine, pick the provider with
--provider. - Installing the wrong package.
npm install -g claude-memdoes not install the plugin, only the library. - Reading
NOASSERTIONas "no license". It means GitHub could not classify it. cmux has two different licenses depending on the folder. - Looking for cmux on Windows or Linux. For now it only exists for macOS.
Frequently asked questions
Which one should I install first?
The one that fixes this week's problem. If your sessions keep losing context, claude-mem; if your interfaces look generic, impeccable; if the agent skips tests, superpowers or agent-skills. One at a time, so you know what changed.
Are they safe?
They are public, active projects, but Anthropic does not verify any of them: Claude Code's own warning says it does not control what plugins include. Safe for you means what you read: license, hooks, MCP servers and where data goes.
Do they work with agents other than Claude Code?
Yes, per their READMEs. superpowers, claude-mem, agent-skills and impeccable document installs for other agents, such as Codex or Gemini. cmux is a terminal, so it works with any agent that runs in it.
What do I do if a plugin does something I did not expect?
Disable it with claude plugin disable or remove it with claude plugin uninstall. If you do not trust the marketplace either, drop it with claude plugin marketplace remove, which uninstalls everything that came from it.
Checklist
- The license is clear; if it says
NOASSERTION, you read theLICENSE. - The repo is not archived and has recent activity.
- You read
hooks/hooks.jsonand know what runs on each event. - You know whether it starts an MCP server or a local service, and which process.
- You know where it sends data, if anywhere, and how to turn that off.
- You installed one at a time and tried it on a project that is not production.
A good plugin saves you hours; one you do not understand costs you them. If you are starting out, install Anthropic's official skills first, which bring no hooks, and review how to give an AI coding agent the right context, because memory and skills only work on top of good context. To understand what a hook is before you accept someone else's, see how a hook protects your .env in Claude Code.
Saw this on Instagram? Comment REPOS on the Reel and I'll send you the guide.
Written by

Published October 30, 2026 · updated November 1, 2026