134 projects · checked 2026-09-20
Built with decision models
What people are actually putting these models inside, five days after the category existed. Sorted by the job, not by star count.
The shape of this list says more than any individual entry does.
Developer tooling dominates, because a coding agent makes hundreds of cheap decisions per session and each one used to cost a full model call. Browser extensions come next, where a decision that takes a second is a decision nobody waits for. Almost nothing here is a chatbot.
109 of the 134 call the hosted API rather than running open weights. That is worth knowing if you are choosing between the two: the ecosystem is currently built on a single vendor, and the open-weight side is younger than the interface. If you would rather not depend on that, start at self-hosting.
By what it does
Developer tooling 35
Code review gates, CI checks, agent harnesses and editor plugins. The densest cluster in the ecosystem, because a coding agent makes hundreds of small decisions per session and each one costs something.
Browser 18
Extensions and automation that judge pages: what to hide, what to click, what is worth reading. Latency matters here in a way it does not on a server.
Open reimplementations 15
People rebuilding the category from open weights: one-pass scorers, constrained decoding over local models, and small models trained to return a calibrated distribution. The fastest-moving cluster here.
Content 10
Moderation, summarisation triage, voice and video. Judging content rather than producing it.
Robotics and games 10
Embodied agents and game-playing systems, where a decision happens every tick and a generated reasoning trace per step is not affordable.
Agents 9
Tool selection, routing and step gating. An agent choosing its next action is the purest version of the problem this class of model exists for.
Search and retrieval 9
Source selection, query routing and relevance filtering. Deciding which retrieved chunks are actually worth passing on is a classic typed decision.
Evals and benchmarks 8
Head-to-head races, task suites and benchmark harnesses. Read the method before the headline: most of these compare on speed and cost, and very few report calibration.
Language integrations 7
Libraries that make a typed decision feel native to a language: pattern matching on an answer in Elixir, a probabilistic if in Ruby, a typed method via BAML.
Trading 5
The highest-stakes cluster here, and the one where calibration matters most. A confidence number you cannot trust is worse than no confidence number.
Databases 4
Extensions that put a typed decision next to the data: classify a Postgres row in SQL, or navigate a graph one choice at a time.
Everything else 4
Projects that do not fit the larger clusters, including data tooling, recruiting and one-off experiments.
Most starred
Star counts read on 2026-09-20. They measure attention, not whether the project works.
By language
Python and TypeScript split the ecosystem almost evenly, which is unusual and tells you this is being adopted by application developers rather than only by ML teams.
Find out what your data actually supports
We turn your historical decisions into a frozen test set, then tell you which model, which threshold, and how much of it you can safely automate.