Affaan Mustafa won the Anthropic x Forum Ventures hackathon by building a full startup in eight hours with Claude Code, then published the exact configuration that did it. The repo is affaan-m/ecc, "Everything Claude Code," and a single thread promoting it has pulled 782,000 views. I went looking for the part where someone who isn't selling it confirms it works. I couldn't find it.
That gap is the whole story. Everything Claude Code might be the best open-source Claude setup on GitHub right now. The thread telling you so is not evidence that it is.
What's actually in the repo
The pitch is that you stop using Claude as one assistant and run it as an orchestrated team. ECC ships a library of skills, dozens of specialized subagents, and ready-made commands, each scoped to one job. A main session delegates to the right specialist when a task calls for it.
The specific subagents are the interesting part. One reviews security against OWASP standards. One manages context so Claude stops forgetting earlier decisions around hour three of a long session. One learns from past sessions so the setup improves with use. Others handle planning, test-driven development, and language-specific review. It's a coordination layer over Claude Code subagents rather than a prompt pack, refined over ten-plus months of daily use.
It's also portable. ECC runs in Claude Code, Cursor, Codex, and OpenCode, on Windows, Mac, and Linux, and it's MIT licensed. That last detail matters more than any benchmark, and I'll come back to it.
The hour-three memory problem is real
The claim I find most credible is the one about memory. Anyone who has run a multi-hour Claude Code session knows the model loses the thread of early architectural decisions as context fills. A subagent whose only job is compacting and re-surfacing those decisions is solving a problem builders actually hit. That's a sign the system grew out of real use.
The OWASP and TDD agents are harder to judge from a description. "Reviews against OWASP standards" is easy to write into a system prompt and hard to do well. Without seeing output, I treat it as a claim, not a feature.
The thread, and the feed around it
Here's what makes me cautious. The announcement ends with "Bookmark this." It's structured for reach, and it got reach. None of that tells you the code is good.
It helps to see the attention economy this lives in. The same week, my feed served a trader giving away "5 lucid 50k DIRECT funded accounts" for a follow, like, repost, and reply. [1] Another account dangled a $4M USDT airdrop with the line "I just watched another person turn $3 into $20k." [2] A hardware brand raffled two controllers for the usual follow-and-repost combo. [3]
“It’s UNDER a month until @MetaWin airdrops us $4M in USDT And I just watched another person turn $3 into $20k… Every single person who deposits on qualifies for part of the airdrop I’ll give away another $500 so 1 of you can go try your luck, comment your wallet Realistically all of you should go deposit before the $4M is dropped You earn a guaranteed airdrop with a chance to hit a 10,000X while y”
I'm not saying ECC is a scam. It isn't; the code is right there, free. I'm saying "the guy gave away his entire playbook for free" is the same rhetorical shape as "I'll give away $500, comment your wallet," and you should notice when a developer tool arrives wearing airdrop grammar. Free and verified are different words.
What builders actually said
Nothing I can cite. In the material I have, there's no substantive review of Everything Claude Code from anyone who has run it, no benchmark, no independent before-and-after. The thread drew 89 replies and not one of them is load-bearing evidence in front of me. That absence is the honest state of play: a strong-looking artifact, a viral pitch, and a verification vacuum.
Before you clone it
The good news is that MIT licensing collapses the cost of finding out for yourself.
- Clone
affaan-m/eccand read the subagent definitions before running anything. The prompts are the product; if they're thin, you'll see it in five minutes. - Test one claim, not all of them. Run a long session with the memory subagent enabled and check whether hour-three recall actually beats your current setup.
- Wait for independent write-ups over the next few weeks. The signal worth trusting is one builder who shipped with Everything Claude Code and shows the diff.
If the repo holds up, it's a free upgrade to how you run Claude Code subagents. Judge it by the code, which you can read today, not by the thread, which was built to be shared.
Sources
- Tweet by @aw_trades_ — @aw_trades_ · community
- Tweet by @GordonGekko — @GordonGekko · community
- Tweet by @GuliKitDesign — @GuliKitDesign · community