Trending GitHub Repos for Claude Code: T3 Code, claude-mem, ECC

By Beau Johnson··8 min read

Quick answer

The five repos that gained the most GitHub stars last week were T3 Code, claude-mem, ECC, Superpowers and TradingAgents. I'd install Superpowers first, try T3 Code if you pay for two or more agents, take only a few pieces of ECC, hold off on claude-mem, and read TradingAgents' code without trading with it.

In this video I count down the five GitHub repos that blew up last week, ranked by the stars they gained on GitHub Trending. Together they picked up more than 33,000 new stars, and most of them upgrade the coding agent you already pay for with a memory, a game plan, or a cockpit. This is part 3 of the series. It's for anyone using Claude Code or Codex who wants to know which of these are worth installing and what to watch out for.

What I cover in this video

  • 0:45 #5 TradingAgents, a pretend trading firm where every employee is an AI agent
  • 3:02 #4 Superpowers, the plugin that makes your agent ask questions and plan before it codes
  • 5:19 #3 ECC (formerly Everything Claude Code): 71 agents, 314 skills, 98 commands and 27 hooks
  • 7:36 #2 claude-mem, memory that carries over between sessions, and its big asterisk
  • 10:39 #1 T3 Code, a free cockpit for Claude Code, Codex, Cursor and more
  • 12:56 Which ones I'd actually install

#5 TradingAgents and #4 Superpowers

TradingAgents (+4,603 stars)

TradingAgents is a free, open-source research tool. You give it a ticker and a date and get back a report and a rating. It does not place real trades. Four analysts run at once, covering company filings, Reddit, the news and the charts. Then a bull and a bear debate those reports, a manager writes the plan, a trader turns it into a proposal, three risk agents (aggressive, neutral and conservative) argue about it, and a portfolio manager makes the final call and logs it. Since mid-October it runs on your ChatGPT Plus or Pro plan instead of an API key, which explains its second wind. To try it, clone the repo and run:

pip install .
tradingagents

The catches: the "not financial advice" link in the README went to a page-not-found page when we checked on November 3. The paper's strong returns come from a three-month backtest in early 2024, and their own screenshot shows 45 model calls for a single decision. What's worth stealing is the pattern: make your agents argue before they decide, use a cheap model for grunt work and a strong model for the final call.

Superpowers (+5,956 stars)

Superpowers is in the official plugin stores of Anthropic, OpenAI and xAI, with more than a million installs from Anthropic's store alone. It fixes agents that jump straight to code with no tests and then tell you they're done. It's 15 skills (just markdown files) plus one hook. The workflow goes brainstorm and design doc, isolated git worktree, a plan of 2-5 minute tasks, a fresh subagent per task with the test written first, then a review after each task. By my rough count it's only about 1,400 tokens always on. To install it in Claude Code:

/plugin install superpowers@claude-plugins-official

Start a fresh session and say "Let's make a React to-do list." If it's working, it asks you questions first. Watch out: it's overkill for small fixes, and a subagent per task burns usage. One user reported a simple plan on Codex ate their whole 5-hour budget on a Plus plan.

#3 ECC and #2 claude-mem

ECC (+6,201 stars)

If Superpowers is a work ethic, ECC is the whole hardware store: one power user's full setup, packaged. A setup wizard lets you pick a hook profile (minimal, standard or strict). Hooks load your context at session start, 10 checks guard tool calls (like blocking destructive shell commands), you work with specialist agents like a planner, code reviewer and security reviewer, and eight more hooks fire when the agent stops, covering formatting, type checks, a session summary and a cost tracker. A background "instincts" agent exists but is off by default.

The menu costs tokens. By my rough count from the files, if every description got listed that's around 30,000 tokens before you type, compared with about 1,400 for Superpowers. Claude Code may trim that list, so check with /context before and after. Also, the README warns that third-party re-uploads may contain malware, so only install from the official repo.

claude-mem (+7,255 stars)

claude-mem gives Claude Code a memory that carries over between sessions. Hooks watch your tool calls, a cheap model (Haiku on your Claude plan by default) writes short notes, those notes are stored locally in SQLite plus a vector database, and your next session starts with 50 of them as a briefing. Your agent can then search over MCP: index, then timeline, then full details.

I think it's the most useful idea on the list, and it has the biggest asterisk. According to the project's own release notes, from October 25 to November 2 a fresh install through the browser setup uploaded the full text of your last 30 days of prompts, with session IDs, project names, folder paths and timestamps, without asking. A fix shipped November 2 that adds a consent screen and lets you keep everything local, but the notes say prompts already uploaded are not removed by that update. Analytics are on unless you opt out. If you still want to try it, their notes tie the upload to the browser setup, not the plugin route. Add their marketplace with /plugin marketplace add, then:

/plugin install claude-mem
npx claude-mem telemetry disable

#1 T3 Code (+9,770 stars)

T3 Code is the one real spike on the list. On October 30 the weekly page showed about 2,000 stars, and by November 3 it was almost 10,000. It's a free, MIT-licensed cockpit for Claude Code, Codex, Cursor, Grok Build and Antigravity, on desktop, web and phone. Every task is a thread with its own git worktree, diff, preview and a queue that survives restarts. A delegate task tool lets Claude plan and hand the build to Codex. You can steer it from your phone over T3 Connect, which is free for up to three machines. It was built by Theo Brown and Julius Marminge at ping.gg.

brew install --cask t3-code

On Windows it's available through winget. Log into Claude Code and Codex, add a project folder, and try planning with Claude and building with Codex. Before you do, read the catches: the README says "very very early ... Expect bugs," there are over 1,600 open issues, new threads default to Full access (change that in settings first), their FAQ says running Claude through the Agent SDK drains your 5-hour limits faster, and usage analytics are on by default.

Key takeaways

  • Superpowers first. It has the smallest always-on cost of the five, and I think it changes how your agent behaves the most.
  • T3 Code if you pay for two or more agents. Use a spare machine first, switch off Full access, and turn off analytics if you care about that.
  • ECC as a cookbook. Take three to five skills or the minimal hooks, and don't stack it on top of Superpowers.
  • claude-mem, not yet. The idea is great, but the fix is brand new. I'd wait for a track record.
  • TradingAgents: read the code, don't trade with it. The multi-agent debate pattern is the real lesson for building AI agents.

FAQ

What is T3 Code?

T3 Code is a free, open-source, MIT-licensed app from ping.gg that runs Claude Code, Codex, Cursor, Grok Build and Antigravity from one window, on desktop, web and phone. It's still very early, so expect bugs, and change the Full access default in settings first.

Is claude-mem safe to install?

According to the project's own release notes, fresh installs through the browser setup between October 25 and November 2 uploaded the last 30 days of prompts without asking. A November 2 fix added a consent screen, but prompts that were already uploaded aren't removed. I'd wait for a track record before installing it.

Should I use Superpowers and ECC together?

I wouldn't stack ECC on top of Superpowers. ECC's descriptions can add around 30,000 tokens by my rough count, compared with about 1,400 for Superpowers, so take only the few ECC skills you need and check with /context.

Does TradingAgents place real trades?

No. It's a research tool that returns a report and a rating for a ticker and date. Its paper's strong returns come from a three-month 2024 backtest, and none of this is financial advice.

Chapters

  1. 0:005 repos that blew up last week
  2. 0:53#5 TradingAgents
  3. 2:51Join Shipping Skool
  4. 3:11#4 Superpowers
  5. 5:26#3 ECC (formerly Everything Claude Code)
  6. 7:54#2 claude-mem
  7. 10:35#1 T3 Code
  8. 13:19Which ones I'd actually install

Transcript

Read the full transcript

0:00These are the five repos that blew up on GitHub last week. And they took the coding agent you already pay for, the one that forgets everything and jumps straight to code, and gave it a memory, a game plan, and a cockpit. Together, more than 33,000 new stars in one week. In this video, we're counting down all five, ranked by the stars they gained last week. I skipped the ones we already covered. So, this is the week agents got upgrades. A free cockpit for every coding agent you pay for. A memory for Claude that remembered a little too much. 314 skills in one install. The plugin that makes your agent ask questions before it writes code. And a fifth one that puts your agents in suits to run a pretend trading desk with a

0:45disclaimer link that goes nowhere. Plus, which ones I'd actually install, and it's not all five. Let's jump right into it. So, number five is trading agents. more than four and a half thousand stars last week and more than 120,000 total. So, what is it? A free open-source pretend trading firm where every employee is an AI agent. You give it a ticker and a date and get back a report and a rating. It's a research tool. It doesn't place real trades. So, the pain is one chatbot, one confident opinion. So, here's how it actually works. Step one, four analysts run at once. company filings, Reddit, the news, and the charts. Step two, a bull and a bear debate those reports and a manager

1:30writes the plan. Step three, a trader turns it into a proposal. Step four, an aggressive, a neutral, and a conservative risk agent argue about it again. And then step five, a portfolio manager makes the call and logs it so later runs can grade it. So, who built it? Almost single-handedly, Yijia Xiao, a UCLA PhD going by his GitHub profile. So why last week? It's a second wind. In midocctober, it started running on your chat GPT plus or pro plan instead of an API key. So to try it, clone the repo, pip install dot, and type trading agents. Pick a date in the past and read the debate, not the rating. The AI lesson is the pattern. Make your agents argue before they decide. Cheap model for grunt work, strong model for the

2:16final call. So here's something to pay attention to. The not financial advice line in their readme links to a page that when we checked on November 3rd said page not found. To be fair, their terms page says it too. The paper claims strong returns but on a 3month back test from early 2024. Their own screenshot shows 45 model calls for one decision. And in that screenshot, after the whole firm argues, the last message in the log says, "If you want, I can also provide a tactical trading recommendation. What do you think?" The hedge fund is still a chatbot. Real quick, every video I make is built in code with Claude Code. Inside Shipping School Premium, I show you how step by step to turn your product into videos that get views.

3:02Plus, new courses every day, 500 plus hours of replays, live calls every week, and a Q&A with me every Saturday. Join free link below. So, number four is superpowers. almost 6,000 stars last week and more than 310,000 total. So why is a repo that big still pulling in 6,000 stars in a week? It's a default now. It's in the official plug-in stores of Anthropic, OpenAI, and XAI with more than a million installs from Anthropic Store alone. And the pain it fixes is agents jumping straight to code with no tests and telling you it's done. So here's how it actually works. It's 15 skills, which are just markdown files, plus a hook that tells your agent you have superpowers. Then it says, "If

3:48there's even a 1% chance a skill might apply, you absolutely must invoke the skill." Absolutely must, in all caps. Step one, brainstorming. It asks you questions and writes a design doc before it builds anything. Step two, an isolated Git work tree. Step three, a plan of tiny tasks, two to five minutes each, clear enough for quote, "An enthusiastic junior engineer with poor taste, no judgment, no project context, and an aversion to testing." That's the agent. Step four, a fresh sub agent per task. And the test comes first. And then step five, a review after every task. By my rough count, it's only about 1,400 tokens always on. So, who built it? Jesse Vincent, who created Request

4:34Tracker, ran the Pearl 5 project, and co-founded Keyio. On the project's first birthday, he wrote about the morning he shipped version one. Quote, "I screwed up that one line of code and then had to scramble to fix it." He says that was the last line of code he wrote. So, to install it in Claude Code, type /pluginstall superpowers at claude plugins official. Start a fresh session and say, "Let's make a React to-do list." If it's working, it asks you questions first. So, it makes you better with AI by taking the project manager job off your plate. So, here's something to pay attention to. It's overkill for small fixes. There's an open issue titled, "Version 7's brainstorm questions are rather annoying, and the project's own notes agree. A sub agent per task also

5:19burns usage." One user reported, "A simple plan on Codeex ate their whole 5-hour budget on a plus plan." So, number three is ECC. more than 6,000 stars last week and more than 295,000 total. It was called Everything Clawed Code until it was renamed ECC in May. So, what is it? If Superpowers is a work ethic, ECC is the whole hardware store. One install, 71 agents, 314 skills, 98 commands, and 27 hooks. So, the pain it fixes is building your own setup from scratch. This is one power user's whole setup, packaged. So, here's how it actually works. Step one, a setup wizard where you pick a hook profile, minimal,

6:05standard, or strict. Step two, hooks load your context when a session starts. Step three, 10 checks guard your tool calls like blocking destructive commands before a shell command runs. Step four, you work with specialist agents, a planner, a code reviewer, a security reviewer. Step five, eight more hooks fire every time the agent stops. Formatting, type checks, a session summary, a cost tracker. And then step six, if you turn it on, a background agent writes what it calls instincts and/ evolve turns them into new skills. So who built it? Afan Mustafa. He says he won an anthropic hackathon in New York in September of last year, and the prize was $15,000 in claude credits. In January, he posted a guide on X. And the

6:50repo went up the next day. By the repo's own Star Log, 16,000 stars in under five days. So why last week? No new release since October 1st. It just kept stacking close to 900 stars a day. So the guided install is NPX ECC universal setup. Pick the minimal hooks and run / context before and after. And here's how it makes you better with AI. It's the biggest public look at how a power user sets up Claude code. So read the skills folder and take the three to five you need. Here's something most people don't think about. The menu costs tokens. By my own rough count from the files, if every description got listed, that's around 30,000 tokens before you type versus about 1,400 for superpowers.

7:36Cloud code may trim that list. So check with /context. And the tagline in the readme, optimize the context window. The install adds 314 skills. Come on now. The readme also warns that third-party re-uploads may contain malware, so stick to the official repo. So, number two is Claude Mem. More than 7,000 stars last week and more than 111,000 total. It's a free plugin that gives Claude code a memory that carries over from one session to the next. I think it's the most useful idea on this list with the biggest asterisk. So, the pain is that every new session starts with amnesia. What did we decide about Oth yesterday? Part two's AI memory went after the same pain locally. So, here's how it actually

8:23works. Step one, hooks watch every tool call your agent makes. Step two, a cheap model haiku on your cloud plan by default writes short notes, a decision, a bug fix, a feature. Step three, the notes get stored on your machine in SQLite plus a vector database. Step four, your next session starts with 50 of those notes as a briefing. And then step five, your agent gets a three-step search over MCP index, then timeline, then full details. So it only pulls what it needs. So who built it? Alex Newman, who goes by the Mac on GitHub. So why last week? A big product push about 30 releases in 4 weeks and a new browser setup on October 25th. And it makes you

9:08better with AI because new sessions start with what you already decided, not a blank slate. So, here's something to pay attention to, and I'm reading it straight from the project's own release notes. From October 25th to November 2nd, a fresh install through the browser setup, quote, "Scanned local agent histories and uploaded the full text of the last 30 days of prompts with session IDs, project names, folder paths, and timestamps without asking." That's their wording, not mine. To be fair, the upload was mentioned in their own change log the day it shipped. It just never asked you on screen. They shipped a fix on November 2nd. Now it stops and asks first with a consent screen and you can keep everything local. Their notes also say prompts already uploaded are not

9:53removed by this update. And the screen that was doing the uploading, it said checking your AI usage. So if you installed it through the browser setup in that window, update and ask cm.ai what they kept. Also, their docs say analytics are on unless you opt out and the note-taking model runs on your plan. So, if you still want to try it, their notes tie the upload to the browser setup. The plug-in route isn't described as affected. In Claude Code, it's /plugin marketplace add/cloudme then/pluginstall claude mem then npxclaude mem telemetry disable. Then open a new session the next day and see what it remembers. So number one on our list is T3 Code.

10:39Almost 10,000 stars last week and more than 43,000 total. This is the one real spike. On October 30th, GitHub showed about 2,000 stars for the week. When I pulled it again on November 3rd, almost 10,000. So what is it? A free open-source cockpit for the coding agents you already pay for. clawed code, codeex, cursor, grock build, and anti-gravity from one window, even on your phone. The readme asks, "Wait, what are you selling me?" Nothing. So, the pain is juggling a terminal per agent. I built Becken, my own agent setup, because I was tired of that exact thing. So, I get it. I get it. So, here's how it actually works. Step one, you log in the agents you already have. Step two,

11:24you run the T3 server or open the desktop app. Step three, every task is a thread with its own git work tree, a diff, a preview, and a queue that survives restarts. Step four, agents manage agents. A delegate task tool lets Claude plan, hand the build to codeex, and get the results back. And then step five, you steer it from your phone over T3 Connect free for up to three machines while the work runs on your desktop or a server. So, who built it? Theo Brown and Julius Marm at ping.gg. and Julius is the top committer by a mile. So why now? The first stable release of what they call Orchestrator V2 came out October 15th. Then on October 29th, Theo posted a video called T3 Code Broke Me and it's

12:10at about 400,000 views. Theo says it's about to hit 550,000 users. The description of the same video says 700,000. Either way, a lot. So on a Mac it's brew install--cask t3-CODE and on Windows it's winding. Log into Claude code and codeex. Add a project folder and try this. Plan it with Claude, build it with codeex. That's how it makes you better with AI. The best model for each step and long runs off your laptop. Here's something to pay attention to. The readme says we are very very early in this project. Expect bugs. There are over 1,600 open issues. New threads default to full access, so change that in settings first. Their FAQ says running Claude through the agent

12:56SDK drains your 5-hour limits faster than using Claude code directly and usage analytics are on by default. And the funniest part, the readme calls it an agent harness control surface. Then the FAQ asks, is T3 Code a harness? No. And when Theo listed the downsides in his own video, he said it's made by Theo, which sucks. So, which ones would I actually install? Superpowers first. By my count, it has the smallest always on cost of the five, and I think it changes how your agent behaves the most. T3 code if you pay for two or more agents. Spare machine first. Switch it off full access and turn the analytics off if you care. ECC, not the whole thing. Read it like a cookbook. take

13:42three to five skills or the minimal hooks and I wouldn't stack it on top of superpowers. Claude mem not yet. Great idea, but the fix is brand new, so I'd wait for a track record. And trading agents, read the code, don't trade with it. So, what's the repo I missed? Drop it in the comments and the best one could end up in part four. And if superpowers or T3 code ends up on your machine, that alone is worth a like. Have a blessed day.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading