Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
The best Claude Code that $200 can buy
| Date | Stars |
|---|---|
| 2026-07-31 | 269 |
| 2026-08-06 | 269 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# nanocode
> The best Claude Code that $200 can buy.
`nanocode` is a library for training your own Claude Code model end-to-end using [Constitutional AI](https://arxiv.org/abs/2212.08073) including tokenizer training, pretraining, synthetic data-generation, agentic SFT with tool use, and DPO with constitutional alignment.
For a detailed writeup, please see the [announcement post](https://github.com/salmanmohammadi/nanocode/discussions/1).
`nanocode` is written in pure JAX and designed to be run on TPUs, and you can get started right away using the [Google TRC program](https://sites.research.google/trc/about/) which gives you free access to TPUs for a month (I think new Google Cloud accounts also get $300 free). You can reproduce nanocode-d24 (1.3B params) in ~9 hours on a TPU v6e-8 for $200, or train nanocode-d20 (~500M params) in ~1.5 hours for $34.
| depth | params | CORE | cost | time | MFU | fwe bpb | sv2 bpb |
|-------|--------|-------|------|---------|-------|---------|---------|
| d12 | 135M | 0.090 | $3 | 9 min | 17.4% | 0.956 | 0.689 |
| d20 | 477M | 0.170 | $34 | 1.4 hrs | 45.2% | 0.838 | 0.533 |
| d24 | 1.3B | 0.227 | $200 | 9.3 hrs | 52.5% | 0.759 | 0.445 |
### Getting started
If you're following along using TPUs, let's first set up [`gcloud`](https://docs.cloud.google.com/sdk/docs/install-sdk) locally to manage our TPU instances:
```bash
# this is a mac-specific installer - adjust as per your system
curl -O https://dl.google.com/dl/cloudsdk/channels/rapid/downloads/google-cloud-cli-darwin-arm.tar.gz
tar -xf google-cloud-cli-darwin-arm.tar.gz
./google-cloud-sdk/install.sh
```
You'll then need to do a bit of administration in setting up your Google Cloud project ([see prerequisites #1 and #2 here](https://docs.cloud.google.com/tpu/docs/managing-tpus-tpu-vm#prerequisites)). Once you've done all that, let's spin up a TPU v6e-8 pod:
```bash
export GCLOUD_ID=YOUR_TPU_PROJECT_ID # from your Google Cloud TPU project setup
gcloud compute tpus tpu-vm create nanocode \
--zone=us-east1-d \
--accelerator-type=v6e-8 \
--version=v2-alpha-tpuv6e \
--network=default \
--project=$GCLOUD_ID
# --spot uncomment this if you're on the TRC program, as only pre-emptable TPU v6 pods can be provisioned
```
Once your pod is available (check with `gcloud compute tpus tpu-vm list --zone=us-east1-d`), SSH in:
```bash
gcloud compute tpus tpu-vm ssh nanocode --zone=us-east1-d
```
I like to develop on my local machine and use a lightweight `entr` process to synchronise changes through `rsync` to the TPU pod:
```bash
# on your local machine, in a separate tmux/screen/TTY window
git clone [email protected]:salmanmohammadi/nanocode.git
# watches for file changes and automatically pushes to your pod
find ./nanocode \
\( -path '*/.git' -o -path '*/.venv' -o -path '*/__pycache__' \) -prune \
-o \( -name '*.py' -o -name '*.toml' -o -name '*.sh' \) -print \
| entr -r rsync -avz \
--exclude='.git' --exclude='.venv' --exclude='__pycache__' \
--exclude='uv.lock' \
./nanocode/ -e "gcloud compute tpus tpu-vm ssh nanocode --zone=us-east1-d --" :~/nanocode/
```
But you can also just `git clone` inside your pod directly. Now let's set our environment up:
```bash
# on your TPU pod
cd nanocode/
./install.sh tpu
source .venv/bin/activate
```
Now you can kick off any of the `speedrun` scripts:
```bash
> ls -1 speedrun_
speedrun_d12.sh # 135M params
speedrun_d20.sh # 477M params
speedrun_d24.sh # 1.3B params
speedrun_d3_cpu.sh # 4M params - mostly for debugging on CPU
speedrun_d6.sh # 23M params
```
<details>
<summary><strong>Multi-slice TPU pods</strong></summary>
For larger accelerator configs (e.g. v6e-32), you'll be working with a multi-worker pod, so the steps are a little different:
```bash
gcloud compute tpus tpu-vm create nanocode \
--zone=europe-west4-a \
--accelerator-type=v6e-32 \
--version=v2-alpha-tpuv6e \
--network=default \
Excerpt of 10,769 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:b65e1129706306cc, llm:description: 'The best Claude Code that $200 can buy' (no topics, README empty). Likely prompts or code for Claude (AI assistant)—appears related to prompt-engineering / ai-coding. No clear model or dataset; likely prompt/code snippets for AI usage.
matched fp:b65e1129706306cc, llm:description: 'The best Claude Code that $200 can buy' (no topics, README empty). Likely prompts or code for Claude (AI assistant)—appears related to prompt-engineering / ai-coding. No clear model or dataset; likely prompt/code snippets for AI usage.