Iris MCP server

Iris is a small helper that watches over your AI agent's work. It saves each run of your agent to a file on your own computer and gives it a score for quality, safety and cost. It is handy if you build or test AI agents and want real numbers instead of just eyeballing the output.

What this MCP server does

You ask your AI to log a task and check the result. Your AI calls Iris, and Iris saves the run into a small database on your computer. Then Iris runs its built-in rules over that run: it looks for private information, sneaky instructions hidden in text, made-up facts, and costs that are too high. You get back a verdict, and you can open a dashboard in your browser to see every run, what failed, and why.

Click to zoom

What you can do with it

• Log a run of your agent and get a quality score
• Check an output for private information like names or emails
• Spot hidden instructions that try to trick your AI
• Catch numbers or facts the source never said
• Watch how much each run costs and set a limit
• Compare two runs on the same set of questions
• See every failure on a dashboard, worst first

Try asking your AI

• “Log that last task to Iris and evaluate the output.”
• “Check this reply for any private information before I send it.”
• “Show me the runs that failed today and why.”
• “Compare the last two runs on the same questions.”

What it gives back to you

You get a verdict for each run, with the state (pass or fail), the reason, and which rules fired. In the chat it looks like a short answer plus the details. On the dashboard you see cards for each failure, naming the rule and the evidence, and lists of runs you can click into.

Before you start

Who it's for

People who build or test AI agents and want clear numbers on quality, safety and cost instead of guessing.