ecodex brings the same discipline to everything outside Claude. It is the OpenAI Codex command line with a per-action firewall and calibration built in, and it runs on any OpenAI-compatible model: GPT, Qwen, DeepSeek, or a local one. The model still has to earn the right to act.
A firewall on every action
State-changing commands are gated on whether the model has actually grounded itself first. Because the gate sits below the model, even a non-GPT model gets the same protection.
Confidence that is measured
Each unit of work is checked against real evidence, so the model's stated confidence is earned rather than asserted. The same calibration empirica brings to Claude Code.
Bring your own model
Point it at any OpenAI-compatible endpoint: a provider, a router, or a local model. No lock-in to one vendor.
The same loop, in your Codex harness
ecodex runs the exact shape empirica does, just below a different model. A plan is broken into small measured chunks with their own goals and tasks. Inside each one the model has to investigate before it acts, and it only earns the right to act once its confidence is grounded in what it actually found, not in a guess. When the chunk closes, the work is measured against real evidence.
See the full shape on the empirica page →empirica is the framework and the discipline. ecodex is one harness that puts it to work, the way in for OpenAI Codex and open-weight models. Same principles, a different coding tool.
Get started
Two parts. ecodex needs the empirica CLI on your PATH for the
discipline (the firewall and the calibration) to work, then ecodex itself.
# 1. The empirica CLI (the discipline)
pip install empirica
# 2. ecodex (the harness)
cargo install --git https://github.com/EmpiricaAI/ecodex codex-cliPrebuilt binaries are on the GitHub releases page. A Homebrew tap is on the way. ecodex is alpha and under active development.