# OpenInterpretability

> Independent lab for AI-agent safety: pre-registered research and open tools. Main product: Ekbasis, an open
> consequence model that forecasts what an action will do before an agent runs it (typed questions, calibrated
> probabilities, one forward pass). Open weights (Apache-2.0) on Hugging Face; you run it yourself. The hosted API at
> https://openinterp.org/api/v1 was discontinued on 2026-10-20 (it answers 410).

## Ekbasis

- [Ekbasis for AI agents](https://openinterp.org/ekbasis/agents.md): exact API formats, when to consult it, decision rules, failure handling. Start here if you are an agent.
- [Getting started (human guide)](https://openinterp.org/ekbasis/start): mental model, writing a good state, real examples by domain.
- [Ekbasis overview](https://openinterp.org/ekbasis): results, builds, papers.
- [Download the model](https://huggingface.co/caiovicentino1/Ekbasis-Skopos-27B-INT4) (the model card has the four-command quick start; also Ekbasis-27B-INT4 and Ekbasis-27B-MLX-4bit).
- [Client, CLI guard, Claude Code hook, MCP server](https://github.com/OpenInterpretability/ekbasis)
- [Cookbook: measured use cases, agent skill, studies](https://github.com/OpenInterpretability/ekbasis-cookbook)
- [Paper: When Does a Consequence Model Make AI Agents Safer?](https://doi.org/10.5281/zenodo.23197341)