/Tech16h ago

Developer Niels launches share-codex and a dataset of 16,000 user-agent interaction turns to train open AI agents

The tool uploads local agent traces directly to Hugging Face.

67765217K

#217

Original post

Dimitris Papailiopoulos@DimitrisPapail#217inTech

This is extremely generous of Niels. I'll consider doing the same.

The agent of the future will continual learn from user interactions and the tasks it has worked on since day 0. There is so much models can learn from their environment. High quality data for free.

Niels Mündler-Sasahara@nielstron

If we want open agents with great world models we need public datasets with massive amounts of agentic interactions. With terminals and with users.

For that reason I am releasing my personal 16k user turns, collected over the last few months.

Alongside I releae share-codex. The tool scrapes your agents local session database, lets you choose which traces to publish, and uploads them to HF. Easy to setup and run!

If you want to try it and check out my traces, links in the comments.

7:55 AM · Jun 18, 2026 · 7.2K Views

Sentiment

Users are excited about the release of 16K agent interaction traces and the Share-Codex tool because they look forward to insights from autoresearch loops and appreciate the open sharing.

Pos

100.0%

Neg

0.0%

2 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

X (FORMERLY TWITTER)Via

#217

Posts from X

Most Activity

VIEWS356LIKES7

Eric W. Tramel@fujikanaeda

@nielstron Niels, try running http://moraine.sh cross all harnesses, gives a database of normalized traces you can export to anywhere. let me know your thoughts, export features for training formats would be a good feature to add in roadmap

13h35673

BOOKMARKS10

Niels Mündler-Sasahara@nielstron

https://huggingface.co/datasets/nmuendler/share-codex

https://github.com/nielstron/share-codex

17h319510

RETWEETS4

Niels Mündler-Sasahara@nielstron

If we want open agents with great world models we need public datasets with massive amounts of agentic interactions. With terminals and with users.

For that reason I am releasing my personal 16k user turns, collected over the last few months.

Alongside I releae share-codex. The tool scrapes your agents local session database, lets you choose which traces to publish, and uploads them to HF. Easy to setup and run!

If you want to try it and check out my traces, links in the comments.

Dimitris Papailiopoulos@DimitrisPapail

http://x.com/i/article/2056344151235387392

17h9.9K3336

Vaish Shrivastava@VaishShrivas

@nielstron Amazing! Thank you for sharing

13h1551

faiaz@faiazAz

@DimitrisPapail Looking forward to seeing your agent traces, particularly your autoresearch loops.

14h59