Agents Md Generator

What is Agents Md Generator?

Agents Md Generator is an MCP server that analyzes codebases using tree-sitter AST parsing and produces AGENTS.md files — a README formatted for AI agents rather than human readers. It is compatible with any MCP-capable client, including Claude Code, Gemini CLI, Cursor, and Windsurf.

How to use Agents Md Generator?

Install with Python 3.11+, uv, and Git. Register the server in your MCP client’s config (e.g., claude mcp add agents-md uvx agents-md-generator). Then ask your AI client: “Generate the AGENTS.md for this project”. The client automatically calls the generate_agents_md tool. Optionally set project_path (defaults to current directory) and force_full_scan (skips cache). Large projects are handled via a chunked streaming tool.

Key features of Agents Md Generator

• Tree-sitter based AST parsing across five languages
• Incremental scanning with SHA-256 change detection
• Chunked streaming of large analysis payloads over MCP
• Cache and temporary files stored outside the project directory
• Support for custom exclusion/inclusion patterns via .agents-config.json
• Environment variable and entry point detection

Use cases of Agents Md Generator

• Generate AGENTS.md for a new project from scratch
• Automatically update AGENTS.md when the codebase changes
• Share consistent project context across different AI coding assistants

FAQ from Agents Md Generator

How does incremental scanning work?

On the first run all git-tracked source files are cached. Subsequent runs only re-parse files whose SHA-256 hash changed. For modified files, only changed public symbols are included. If nothing changed, the tool asks whether to improve the existing AGENTS.md anyway.

What languages are supported?

Python, C#, TypeScript, JavaScript, and Go.

Does it modify my project files?

No. All cache and temporary payload files are written to ~/.cache/agents-md-generator/<project-hash>. The only file written to your repository is AGENTS.md, and that is created by your AI client (not by the server directly).

Can I customize what gets analyzed?

Yes. Create an optional .agents-config.json in the project root. You can set exclude and include globs, restrict languages, change the output path, and adjust the impact_threshold (high/medium/low) that controls which changes trigger regeneration.

Is it compatible with any MCP client?

Yes. It uses stdio transport, so any client that supports the mcpServers configuration can use it. Examples include Claude Code, Gemini CLI, Cursor, and Windsurf.