graphdb

A Kùzu-backed code-graph MCP plugin for Claude Code. Indexes Ruby on Rails codebases into a queryable graph of associations, routes, callbacks, jobs, and mailers — a token-efficient alternative to grep for navigating Rails code.

The animation above shows the CLI. The same eight tools are exposed to Claude Code via MCP — see docs/PROMPTS.md for natural-language prompts you can try directly in a Claude Code session.

Why

Rails encodes most of its architecture in DSL declarations (has_many, before_action, resources, include) rather than direct method calls. Grep matches the strings but cannot turn them into relationships:

Related MCP server: sonde

Status

Proof of concept, validated against a real Rails 8 / Ruby 3.3 codebase (smart-hub-backend, 494 indexed .rb files):

nodes: 2913    edges: 4731    index time: ~0.8s
60 models   63 controllers   25 concerns   121 services   67 serializers
7 mailers   6 jobs           109 migrations   230 routes   421 actions

Validation against grep ground truth on smart-hub-backend:

Quickstart

# Set up a venv with uv (sidesteps Homebrew's broken ensurepip on macOS)
uv venv .venv --python 3.13
uv pip install --python .venv/bin/python -e .

# Index a Rails project (writes to ~/.cache/graphdb/<repo>-<hash>.kuzu/;
# nothing is written to the target repo)
.venv/bin/graphdb --root /path/to/rails_app index /path/to/rails_app

# CLI queries
.venv/bin/graphdb --root /path/to/rails_app query def User
.venv/bin/graphdb --root /path/to/rails_app query associations User
.venv/bin/graphdb --root /path/to/rails_app query routes Api::V1::UsersController
.venv/bin/graphdb --root /path/to/rails_app query callers UserInviteMailer --depth 2
.venv/bin/graphdb --root /path/to/rails_app query impact ScheduleExecutionJob
.venv/bin/graphdb --root /path/to/rails_app query refs Account
.venv/bin/graphdb --root /path/to/rails_app stats

MCP integration with Claude Code

Add to your Claude Code MCP config (~/.claude.json or project-level):

{
  "mcpServers": {
    "graphdb": {
      "command": "/path/to/Graph_DB/.venv/bin/graphdb-mcp",
      "env": { "GRAPHDB_ROOT": "/path/to/your/rails_app" }
    }
  }
}

Tools exposed to Claude:

Graph schema

NODE KINDS
  file, module, class, method, class_method,
  model, controller, action, concern, service, serializer,
  job, mailer, validator, helper, channel, error_class,
  migration, route, external

EDGE KINDS
  Generic Ruby:    CONTAINS, CALLS, INHERITS, INCLUDES
  Model DSL:       HAS_MANY, BELONGS_TO, HAS_ONE, HABTM, VALIDATES, SCOPES
  Controllers:     BEFORE_ACTION, AFTER_ACTION, SKIP_BEFORE, RENDERS
  Routes:          HANDLES, MOUNTS
  Async:           ENQUEUES, DELIVERS

How it works

1. indexer.py parses every .rb under app/, lib/, config/, db/migrate/ with tree-sitter-ruby. Two passes:

   • Pass 1: register every class/module/method as a node, build a qualname index used for resolution.

   • Pass 2: walk again, emit INHERITS, INCLUDES, CALLS, and the Rails-specific edges (HAS_MANY, BEFORE_ACTION, ENQUEUES, etc.)

   • Pass 3: a separate walker handles config/routes.rb (DSL outside any class), producing Route nodes with HANDLES edges to controller actions.

2. graph.py buffers nodes/edges in memory and bulk-flushes via Cypher UNWIND at end of indexing. Drops full-repo indexing from ~8 minutes (individual inserts) to ~0.8 seconds.

3. queries.py issues Cypher patterns like MATCH (caller)-[:CALLS|ENQUEUES|DELIVERS*1..3]->(target) — the database does the BFS.

4. mcp_server.py exposes the queries as MCP tools for Claude Code.

Known PoC limitations

• Static parsing only. Doesn't run Rails — no type inference, no resolution of define_method / method_missing / dynamic class_eval.

• Relative imports / autoload constants: resolved by unique-name fallback; ambiguous names are skipped silently.

• Custom DSLs (acts_as_paranoid, paper_trail's has_paper_trail, amoeba): not yet specialized — they emit generic CALLS edges at class scope.

• Routes: scope, member/collection custom routes, constraints, concerns (route concerns), nested resources past 1 level — partial or skipped. The 7 standard REST actions, namespace, mount, devise_for, and get/post/... with to: work fully.

• Reindex is one-shot. Edit a file, rerun graphdb index. No file watcher.

Roadmap

• Specialize the major Gemfile DSLs in smart-hub-backend (paper_trail, discard, amoeba, flipper).

• Member / collection / nested resource routes.

• Incremental reindex via file-watcher.

• AI-generated per-method summaries stored on nodes (read summary first, file second).

• Benchmark harness: grep-only vs graph-only Claude sessions on identical tasks.