sdl

SDL-MCP is a code context MCP server that delivers token-efficient, symbol-based context about your codebase to AI agents. Instead of flooding the AI with entire files, SDL-MCP exposes structured symbol cards, dependency graphs, and incremental deltas that can be assembled into task-specific graph slices. This approach minimizes token usage while preserving relevant context and governance through policy and auditing. The server supports a suite of tools organized around the Iris Gate Ladder, enabling progressively richer context as needed: starting with Symbol Cards that summarize symbols and dependencies, moving to Skeleton IR to understand structure, then Hot-Path Excerpts for usage-accurate snippets, and finally Raw Code Windows when full source access is justified. In practice, you connect to SDL-MCP with a compatible MCP client (such as Claude Code or Claude Desktop) and request context tailored to a task, receiving a compact, graph-based representation rather than raw files.

Prerequisites:

• Node.js and npm installed on your system
• Network access to install npm packages

Installation steps:

1. Install the SDL-MCP CLI globally

   npm install -g sdl-mcp
   

2. Verify installation / get the version

   npx --yes sdl-mcp@latest version
   

3. Initialize configuration and graph database (per project)

   sdl-mcp init          # Create config + graph DB path
   sdl-mcp index         # Index your repositories
   

4. Start the MCP server (example using stdio or HTTP if supported by your setup)

   sdl-mcp serve
   

Notes:

• You can also run without a global install using npx directly:

  npx --yes sdl-mcp@latest version
  

• For automated environments, you can invoke via the mcp_config JSON described above to spawn the server programmatically.

Tips and considerations:

• The server exposes 24 tools for code navigation and analysis (example tools include sdl.symbol.getCard, sdl.code.getSkeleton, sdl.code.getHotPath, sdl.code.needWindow, among others). Use the Iris Gate Ladder to request context in increasing detail as needed.
• Context is delivered as symbol cards and graph slices, which keeps token usage low. When full source access is required, requests can escalate to raw code windows under policy governance.
• Ensure your indexing step is completed (sdl-mcp index) before starting the server so the AI clients have a populated graph to query from.
• Policy and auditing controls are available to govern raw code access; always justify code window requests with identifiers and line bounds.
• If you plan to run in a container or cloud environment, you can adapt the mcp_config to use the appropriate command runner and environment variables as needed.