Skip to main content

How to Work With AI Agents

:::note What is this guide about? This guide shows how to point an AI agent, such as a coding assistant in your IDE, at meshStack, so it can create and manage building block definitions and building blocks for you. :::

How Your Agent Learns About meshStack​

meshStack publishes one entry file for AI agents: https://docs.meshcloud.io/llms.txt. From there, the agent finds the agent context, a set of short pages on access, guardrails, concepts and tasks, and three skills with the exact commands:

The agent needs no knowledge of meshStack beyond this link.

Prerequisites​

  • The meshStack CLI installed where your agent runs.
  • A role in the workspace your agent works in. To create building block definitions, you need the Workspace Owner or Workspace Manager role there.

Step by Step Guide​

1. Point Your Agent at meshStack​

Add a few lines to the instructions file your agent reads in your repository, for example AGENTS.md:

## meshStack

Before you work with meshStack, read https://docs.meshcloud.io/llms.txt.
Our meshPanel is at <your meshPanel URL>. Work in the workspace `<workspace identifier>`.

Your agent then reads the agent context when a task needs meshStack. It finds the API URL of your meshStack in the meshPanel page. If it does not find it there, add the API URL to these lines: the URL you pass to meshstack login --endpoint.

If your agent supports skills, you can also add the three skills above to it, so it loads them without a link.

2. Let Your Agent Log In as You​

Your agent works with your own meshStack CLI login. It starts the login, and you finish the browser step. Your agent then acts with exactly your permissions, every action shows up as yours in the event log, and you can revoke the login at any time under Profile > CLI Logins.

tip

An agent that runs without you, for example in a CI/CD pipeline, can't finish a browser login. Give it a workspace API key with only the permissions its task needs, and keep the Key Secret in a secret store.

3. Review What Your Agent Does​

The agent context tells your agent to:

  • show you the Terraform plan before every change, and apply only after you agree,
  • ask you before it deletes or replaces anything,
  • leave run approvals to you: it reports the plan of a run that waits for approval and links you to it in meshPanel,
  • release a building block definition only when you say so.

Concepts​

Guides​