Run the OpenCaliber console

This public website hosts documentation and downloads. Evaluations run in your own Python environment; this site does not collect or evaluate your traces.

Download and start

Download console package

  1. Install Python 3.10 or newer.
  2. Extract the downloaded ZIP into a new folder.
  3. On Windows, double-click start-windows.cmd. Or open a terminal in the extracted folder and run:
python run.py

On macOS or Linux, use python3 run.py.

Open http://127.0.0.1:8787/console on the same computer. No provider key is needed for built-in examples.

Connect an agent

In the running console, choose Connect agents for Python, Node and HTTP examples. Submit completed tool calls or a normalized trace, then inspect agent identity, flagged actions and suggested corrections.

How evaluation works

The engine checks declared actions against configured rules. A violation identifies a triggered rule; review marks uncertainty; clear means no implemented rule triggered. Scores are heuristic severity, not probabilities. It does not execute or contain actions.

Evaluation history stays in the extracted package directory. Keep the desktop server bound to loopback; hosted use requires the separate authenticated HTTPS deployment described in the repository.

Licensing details · Apache 2.0 license