Skip to main content
Once MCP is connected, you can talk to your decisions in plain language. Start by reading, which changes nothing.

​
What you can ask

Claude uses list_models. For each decision you see the current version, owner, your role, and whether any saved run is out of date because the model changed after it.
Claude uses get_model. It reads the decision, the options, every input with its fitted range and plain-language read-back, the calculations, the evidence and the runs.
Claude uses get_influence_diagram, which is generated from the same definition the engine calculates with. It lists the decision, uncertain inputs, calculated values, outcomes and the formula on each relationship.
Claude uses validate_model. It reports missing estimates, shares that exceed 100%, formula loops, inconsistent range answers and inputs still marked not yet known.

​
A good first session

  1. “List my decisions and tell me which have out-of-date runs.”
  2. “Open the first one and explain it in five sentences.”
  3. “Validate it. What is missing before it can run?”
Ask for a specific version if you need one (“version 1 of this model”). Without it, Claude reads the current version.
These requests are read-only, so even a viewer can do all of them.
Next: Build or change a model with Claude