MCP Scorecard

Mission StatementGitHub
← All posts

Four Guardrails for Autonomous Agents: Context, Norms, Think, Undo

A single developer published four MCP servers that form a safety toolkit for autonomous AI agents — check the calendar before acting, verify values are normal, run pre-flight safety checks, and confirm actions are reversible.
net.agentutil

Most MCP servers give agents new capabilities. agentutil gives agents new constraints.

Craig Whitfield published four servers on February 28, 2026 — all MIT-licensed TypeScript, all minimal dependencies (just the MCP SDK and Zod), all sharing the same architecture. Together they form a coherent thesis about what autonomous agents need before they act.

The Four Servers

ServerPurposeExample
context-mcpSituational awareness: holidays, business hours, platform status"Don't send that marketing email — it's Christmas Day."
norm-mcpStatistical baselines: check if values are normal or anomalous"This invoice amount is 14x the category average — flag it."
think-mcpIntent security: pre-flight checks before executing actions"You're about to delete a production database — are you sure?"
undo-mcpReversibility intelligence: can this action be undone?"Sending this email cannot be reversed. Proceed?"

The Thesis

The value proposition is not what each server does individually — it's the category they collectively represent. As AI agents gain more autonomy, the question shifts from "can the agent do this?" to "should the agent do this right now?" These servers provide the pre-flight checklist:

  1. Timing — Is now the right moment? (context-mcp)
  2. Normality — Is this value expected? (norm-mcp)
  3. Safety — Is this action safe to execute? (think-mcp)
  4. Reversibility — Can we undo this if it goes wrong? (undo-mcp)

Each server has its own website (context.agentutil.net, norm.agentutil.net, etc.) and is published separately, suggesting they are designed to be adopted incrementally rather than as a monolith.

What to Know

All four servers score 32 — low, primarily because they are brand new with zero stars and minimal GitHub history. The scores reflect maturity, not concept quality. The idea of agent guardrails as composable MCP primitives is sound and currently underexplored in the registry. If autonomous agents become more prevalent, this category of "safety infrastructure" servers will matter more than most capability servers.

MIT license. All four servers.

Sources: agentutil — context · norm · think · undo · Scorecard: net.agentutil (score 32)

← Code Pathfinder: Security Scanning Through Code Graphs, Not GrepTamper-Proof Audit Trails for the EU AI Act — Five Months Out →