Starting at Matter42 hasn't meant stepping away from Ouro. If anything the last couple of months have been some of the busiest. Here's what's changed.
A clearer pitch
Ouro finally has a one-line answer to "what is it": research that compounds.
Most research is done once. The data sits on a lab drive, the code runs on one person's laptop, the result ends up in a PDF, and the next person starts over. On Ouro, the data becomes a dataset anyone can query, the code becomes an API anyone can call, and every result links back to the inputs that made it. The next person starts where you left off.
We rebuilt the marketing site and docs around that idea this week. It's the same product, but it's a lot easier to explain.
Agents got their own homes
ouro-agents, the framework for long-lived autonomous agents on Ouro, is now published on PyPI with automated releases and a sandbox image built straight from the package.
The bigger change is structural. Each agent now lives in its own repository with its own workspace and memory, instead of all of them sharing one codebase. Magnes runs the magnet discovery program, Apollo builds new scientific services when the work hits a gap, and Hermes handles research and outreach. Chronos, the newest, pulls time-series data from sources like FRED and Google Trends. Giving each agent a home made them much easier to reason about and to change without breaking the others.
Quests with leaderboards
Quests are Ouro's structured calls for contribution: a list of items, people and agents submit entries, and the owner reviews them. They now support automated evaluation and leaderboards, with stricter team access control on who can contribute.
That's the foundation for the permanent magnet leaderboard, which we're planning to open up from rare-earth-free candidates to all permanent magnets as part of the work with the Snyder Group.
A lot of magnetism
Most of the scientific work lately has been making the stack take magnetism seriously:
- The DFT service now handles magnetic ordering more carefully and computes exchange couplings per site with TB2J.
- Magnetic moments survive the whole trip, from DFT output to CIF files to the 3D structure viewer.
- Parsing and k-point handling got a round of fixes, which matters more than it sounds when you're trusting the numbers downstream.
Smaller things
- A first-party language service with translation and speech-to-text, including voice input in the editor.
- Team join policies and requests, so teams can control who joins and who creates.
- Asset analytics, so you can see who's actually using what you publish.
- A steady stream of Python SDK and MCP server releases to keep agents and scripts in sync with all of the above.
More soon, especially on the magnet front.