Agents-md – Scale AI agent context with composable Markdown fragments (github.com)

agents-md is a new tool to scale and manage AGENTS.md context by treating agent context as composable Markdown fragments. Instead of a single static AGENTS.md per repo, agents-md discovers fragment files (default globs '**/agents-md/**/*.md' and '**/*.agents.md'), composes them into deterministic AGENTS.md outputs (one per target directory by default, "nearest" wins), and generates annotated, auto‑owned files you should not edit by hand. Installable via Bun/npm/pnpm/yarn and invoked with init/compose/watch/report (e.g., bun agents-md init; bun agents-md compose; agents-md report --json), it automates keeping agent context current, shareable with human docs, and ready for toolchains that consume agent prompts or context. Technically, agents-md supports in‑fragment directives for routing and imports (<!-- agents-md: target=../docs -->, <!-- agents-md: import=@./shared/common.md -->), priority-based ordering, optional source annotations, truncation and limit policies, and token/size estimates in reports for CI. Configuration is via agents-md.config.(ts|js|mjs) with sensible defaults but full customization of include/exclude globs, per-file filters, and output behavior. It also offers integrations (CLAUDE.md import for tools that don’t read AGENTS.md), a plugin kit for extracting JSDoc/TS types/schemas, and watch mode for hot reload—helping teams build deterministic, reusable, and size‑aware agent contexts across large codebases.