mcp

Andromity features native support for the Model Context Protocol (MCP), allowing your AI agent to interact with external databases, APIs, search engines, and local developer systems.

A common limitation of AI agents with 50+ connected tools is context window bloating. Injecting full JSON schemas for every tool wastes tokens and causes LLM confusion.

Andromity addresses this with Smart Lazy-Loading:

1. Schema Indexing: On startup, Andromity indexes your configured MCP servers and caches simple, compact summaries of what each server's tools do.
2. On-Demand Loading: Full tool schemas are only injected into the active prompt context if the LLM explicitly requests information about that tool or displays intent to call it.
3. Efficiency: Reduces tool-related token overhead by up to 85%, preserving context space for actual code files.

You can manage MCP servers via .andromity/mcp.json (global) or a project-specific .vscode/mcp.json or mcp.json in your repository.

Example configuration:

{
  "mcpServers": {
    "sqlite-db": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sqlite", "--db", "project.db"]
    },
    "github-api": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "github_pat_..."
      }
    }
  }
}

Run the /mcp command in the TUI to view loaded servers, verify connection states, and check available tool names.