rpg-encoder — Claude Code MCP server: a code graph

What rpg-encoder does

rpg-encoder parses a repository with Tree-sitter and pulls out functions, classes and methods along with the edges between them: imports, calls, inheritance. An LLM then reads each entity and writes verb-object features — things like validating JWT tokens or serializing config to disk. Those features cluster into a three-level hierarchy (area → category → subcategory) that follows what the code does rather than the folder tree.

The finished graph compresses into a snapshot of roughly 25K tokens: the hierarchy, entities with their features, a condensed dependency skeleton, and the ten most-connected nodes. The agent reads it once and can then answer questions such as what handles authentication, even when nothing in the code is named auth. The server watches the working tree and re-syncs the graph before the next query, so you never refresh it by hand.

Lifting features runs two ways. Agent lifting spends tokens from your Claude or Cursor subscription. Autonomous auto_lift calls an external model directly and needs an API key in an environment variable, for example ANTHROPIC_API_KEY. On large repos lifting_status tells the agent to hand the loop to a sub-agent or a cheaper model, since feature extraction is pattern matching rather than reasoning. No sub-agent mechanism in your runtime? Run rpg-encoder lift --provider anthropic|openai from the terminal, then call reload_rpg to load the updated graph.