Skip to content
Lupid

Install with an AI coding agent

If you’re driving a coding agent — Claude Code, Cursor, Cline, GitHub Copilot Workspace, or any other LLM-based dev assistant — the installer ships a machine-consumable spec the agent can follow directly. You don’t need to translate the README for it.

The prompt

Section titled “The prompt”

Paste this into your agent of choice:

Install Lupid on this machine by following the instructions at https://raw.githubusercontent.com/akashsebastian333/agentum/opensource/installation/AGENTS.md. Run the install non-interactively, verify the dashboard responds, then print the admin credentials it generated.

The agent will:

  1. Fetch installation/AGENTS.md — a file written explicitly for AI agents to consume.
  2. Detect the OS and run the right platform-specific one-liner with --yes / LUPID_YES=1 so it stays non-interactive.
  3. Run the two curl verification probes documented in AGENTS.md.
  4. Read <install-dir>/.env and surface the admin email and admin password.

What’s in installation/

Section titled “What’s in installation/”

The repo ships a dedicated installation/ directory with four files an agent (or a human) can read:

FilePurpose
install.shBash installer for Linux / macOS / Git-Bash on Windows
install.ps1PowerShell installer for native Windows
AGENTS.mdDecision tree, exact commands, verification probes, recovery table
manifest.jsonSame content as AGENTS.md in structured JSON for tools that prefer it
README.mdHuman-readable overview of the directory

Why AGENTS.md works

Section titled “Why AGENTS.md works”

A coding agent is good at translating intent to commands, but it loops badly on unknown failures. AGENTS.md removes the ambiguity:

  • Decision tree removes “should I try X or Y first?” — the agent reads the right branch based on uname -s.
  • Exact commands with --yes already applied — no interactive-prompt trap.
  • Verification probes with the exact expected JSON response ({"status":"ok",…}) let the agent decide “done” versus “broken” without guessing.
  • Per-symptom recovery table maps every installer error message to the next action, so the agent doesn’t loop trying random fixes.
  • Hard constraints (don’t echo secrets to long-lived buffers, don’t commit .env) stop the agent from doing dumb things.

The structured manifest

Section titled “The structured manifest”

If your agent or pipeline prefers structured input over prose, point it at installation/manifest.json:

{
  "name": "lupid",
  "description": "Runtime security and observability layer for AI agents.",
  "platforms": {
    "linux":   { "command": "curl -fsSL .../install.sh | bash -s -- --yes" },
    "macos":   { "command": "curl -fsSL .../install.sh | bash -s -- --yes" },
    "windows": { "command": "$env:LUPID_YES='1'; iwr -useb .../install.ps1 | iex" }
  },
  "prerequisites": [
    { "name": "docker",         "min_version": "24.0", "check": "docker --version" },
    { "name": "docker-compose", "min_version": "2.0",  "check": "docker compose version" }
    // ...
  ],
  "verify": {
    "dashboard":  "curl -fsSL -m 5 http://localhost:3000",
    "api_health": "curl -fsSL -m 5 http://localhost:3000/api/v1/health"
  },
  "credentials_location": "<install-dir>/.env",
  "credentials_keys": ["AGENTUM_ADMIN_EMAIL", "AGENTUM_ADMIN_PASSWORD"]
}

Constraints for any agent

Section titled “Constraints for any agent”