The Agent Leash

A one-page safety checklist for solo builders running AI agents, coding agents and MCP servers. Run it once today. Run it again before you give any agent a new tool.

Download the PDF

Keys and secrets

  1. Keep MCP configs and .env files out of git.A config file often holds live keys, and a public or leaked repo hands them to anyone.
  2. Give each agent its own scoped key.When one key is shared, one mistake or leak puts every project at risk.
  3. Set a spend cap on every provider key.A loop or a stolen key can burn a month of budget in an afternoon if nothing stops it. See spending limits on AI agents.
  4. Rotate keys on a schedule, and right away after any scare.An old key you forgot about is the easiest one for someone else to find.

Permissions

  1. Give the agent only the folders and tools the job needs.Least privilege means a bad step can only break a small thing.
  2. Start read-only. Add write access one tool at a time.Most jobs begin with reading, and you can always grant more later.
  3. Require your approval for every send, delete, payment and publish.These actions are hard or impossible to undo, so a human should press the button.
  4. Vet each MCP server before you install it.Check who made it, pin the version, and read the tool list so you know what it can do.

Prompt injection

  1. Treat every web page, email and file the agent reads as untrusted data, never as orders.Anyone can hide instructions in text, and the agent may follow them as if you wrote them.
  2. Never let one agent both read untrusted text and hold send or pay rights.Split reading from acting, so a poisoned page cannot reach your money or your inbox.

Logging and telemetry

  1. Check what your logs and telemetry export, and where they send it.If you turn on OpenTelemetry or tool logging, prompts, file paths and secrets can end up in someone else's system.
  2. Keep a plain log of every action the agent takes.When something goes wrong, you need to see what happened without guessing.

Money and control

  1. Set a hard daily spending cap, and make it stop the agent, not just warn you.An alert you read tomorrow does not help with a bill you got today. See how to cap what an AI agent can spend.
  2. Know your kill switch. You should be able to stop everything in under a minute.Write down how to pause the agent, revoke its keys and cut its network access before you need it.

Before it goes out

  1. Read everything yourself before it publishes, ships or sends.The agent does not carry the blame for what goes out under your name. You do.