Agents.md

What is Agents.md?

Agents.md is an MCP server that provides hierarchical context management for AI agents working on long-term projects. It stores project-wide knowledge in AGENTS.md and task-wide scratchpads, enabling persistent context across sessions and subagent isolation.

How to use Agents.md?

Install via the automatic curl script or clone the repository and run with Node or Docker. Configure required and optional environment variables (e.g., HOST, PORT, external AI provider keys). Connect any MCP client to the server’s /mcp endpoint using a user API key.

Key features of Agents.md

• Hierarchical context management (project-wide, task-wide, subagent isolation)
• Persistent knowledge across multiple chat sessions
• Supports external AI providers for subagent tools
• User and API‑key authentication for MCP endpoints
• Admin API for user creation, listing, and key rotation
• Runs locally or in Docker with persistent storage

Use cases of Agents.md

• Long-term project management where an AI agent must retain knowledge
• Multi‑session research tasks requiring accumulated context
• Collaborative workflows with subagents operating on focused context

FAQ from Agents.md

What is Agents.md for?

It is a hierarchical context management system for AI agents working on long-term projects. It stores project-wide and task-wide context, isolates subagents to prevent overload, and persists knowledge across sessions.

How do I install Agents.md?

You can use the automatic install script (Unix-like) or manually clone the repository and run with Node or Docker. See the “Automatic Install” and “Manual Install” sections in the README for detailed commands.

Does Agents.md require an external AI API?

No. Using external AI (e.g., Google, OpenAI) is optional. Set USE_EXTERNAL_AI=true and provide API keys only if you want subagent tools to use an external model.

Where is data persisted?

For Docker deployments, data is stored in the mounted volume (default $HOME/.config/mcp-http-agent-md/data). For Node-based runs, the README does not specify a default location.

How does authentication work?

MCP endpoints require a user API key supplied via ?apiKey= query parameter or Authorization: Bearer header. Admin endpoints use a separate MAIN_API_KEY in the Authorization header. Users are created through the admin API.