MCP file storage for AI agents — hosted markdown, not DIY glue

Model Context Protocol lets agents call tools. ReadmeDB is the hosted MCP server that turns that into a shared project brain: versioned .md files Cursor, Claude Code, Windsurf, and Codex already know how to use.

Why “MCP memory” searches land here

Most MCP memory demos are local SQLite or vector bags. Coding teams need something different: files agents can list, search, diff, and hand to a teammate — the same job as a living CLAUDE.md ecosystem, but cloud-shared across tools.

Endpoint

POST/GET https://app.readmedb.com/api/mcp
Authorization: Bearer rdb_your_key
Protocol: MCP over HTTP (JSON-RPC 2.0)

What agents get (23 tools)

• read_file / write_file — Core durable storage
• search_files — Full-text over the brain
• append_to_file / patch — Safe incremental updates
• get_file_history / diff — Versioned decisions
• get_namespace_summary — Orient at session start
• bulk reads — Fast agent warmup

Client config

{
  "mcpServers": {
    "readmedb": {
      "url": "https://app.readmedb.com/api/mcp",
      "headers": {
        "Authorization": "Bearer rdb_your_key"
      }
    }
  }
}

OAuth discovery is also published for Claude.ai-style clients (/.well-known/oauth-authorization-server). Humans can use the rdb CLI for the same store.

Evaluate MCP storage servers

• Hosted vs local-only (cloud agents need hosted)
• Files you can open vs opaque embeddings
• Version history when an agent overwrites a decision
• Same namespace across Cursor and Claude without sync scripts