I wanted AI agents and developers to share a single source of rules that adapted to context. - A local dev might have bd installed. - A cloud coding agent might have none of them with other resistrictions.

Static AGENTS.md files don’t scale to those realities. They’re either overly general or wrong for half your environments.

How it works

Templates live in .cAGENTS/templates/ with YAML headers that define conditions:

  ---
  name: beads
  when:
    use_beads: "true"
  order: 20
  ---
  ## Issue Tracking with bd
  
  **IMPORTANT:** This project uses bd for ALL issue tracking.
  
  - Check ready work: `bd ready --json`
  - Create issue: `bd create "Title" -t bug -p 1`

  [variables.command]
  use_beads = "command -v bd >/dev/null 2>&1 && echo true || echo false"

When you run cagents build, it evaluates conditions and composes AGENTS.md (and CLAUDE.md, etc.) using only relevant templates.

  If bd is installed → the section appears.
  If not → it’s skipped. Same templates, different contexts.

My goal is to have this library help me implement progressive disclosure for AI agents. I want to present the right amount of context at the right time. I'm experimenting with `cagent context` so an AI agent can look up all relevant rules for any file automatically and On-demand detail granularity. The dream is that agents that don’t need a 300-line AGENTS.md. just the parts that matter for what they’re touching.

How I use cAGENTS:

  - Gitignore instruction files; commit only templates.
  - Auto-build on postinstall.
  - Keep templates small and composable.