Developers · early access

A first verdict in under five minutes

Describe your agent's rules in plain English. An LLM writes the model, one MCP server validates it, and a second checks your agent's choices against it.

1. Install

uv tool install pdtk-val-mcp

Installs the command predictika-validator-mcp. Add it as an MCP server in your client, for example Claude Desktop.

2. Describe your rules

"Support agents can refund up to $250. A team's refunds stop at its monthly budget. Anything over needs a supervisor." Your LLM turns that into a validator model.

3. Point your agent at it

Before each risky action the agent sends its choices. Each message list is written for a different reader: the agent, the audit trail, and the person.

Tools the server provides

Tool	What it does
launch_validator	Starts a validator from a model
validate_model	Checks that a model is well formed
validate_choices	Checks an agent's choices and returns the verdict
reset_session	Starts a fresh scenario on the same validator
delete_validator	Removes a validator you are done with

Example response (illustrative)

{
  "valid": true,
  "llm_messages":   ["Refund exceeds team budget. Ask a supervisor."],
  "audit_messages": ["TeamRefundBudgetExceeded: $4,870 + $180 > $5,000"],
  "user_messages":  ["Your refund needs a supervisor's approval."],
  "assigned": { "amount": 180.0, "team_budget_used": 4870.0 }
}

A soft finding keeps valid true and returns messages. A hard rule failure rejects the choice.

Samples

Start from something that runs

EU flight-delay bot

A conversational agent whose compensation answers are checked against the regulation.

Try the bot

River Crossing puzzle

A four-constraint puzzle that shows why a model can know the rules and still break them.

Read the paper

Everyday validators

Trip check, coupon check, PC build, mortgage program, travel expense and more.

Browse the gallery

Join the early access list Browse the samples