Code Graph Context

CodeGraphContext Skill

Structural code graph queries using the CodeGraphContext MCP server (tree-sitter AST + KuzuDB property graph).

When to Use vs Other Search Tools

Use this skill when you need relationship traversal across the call graph or import graph, not text matching.

MCP Tool Reference

Setup

# Install CodeGraphContext MCP server
npm install -g @codetiger/code-graph-context-mcp

# Index your codebase (run from project root)
code-graph-context index --root . --lang typescript,javascript

Add to .claude/settings.json under mcpServers:

"CodeGraphContext": {
  "command": "code-graph-context-mcp",
  "args": ["--db", ".claude/context/data/code-graph.kuzu"]
}

Usage Pattern

// 1. Find all callers of a function
mcp__CodeGraphContext__find_callers({ symbol: 'shouldUseWorktree', depth: 2 });

// 2. Trace what a function depends on
mcp__CodeGraphContext__find_callees({ symbol: 'routeRequest', file: 'routing-table.cjs' });

// 3. Dead code candidates (low confidence = more results)
mcp__CodeGraphContext__find_dead_code({ scope: 'src/', min_confidence: 0.8 });

// 4. Raw Cypher for custom traversal
mcp__CodeGraphContext__query_graph({
  cypher: 'MATCH (a:Function)-[:CALLS]->(b:Function) WHERE b.name = $name RETURN a',
  params: { name: 'handleAuth' },
});

Integration with Agent-Studio Memory

After graph analysis, record structural findings:

MemoryRecord({
  type: 'pattern',
  content: 'routeRequest has 12 callers — high-risk refactor target',
  area: 'architecture',
});
MemoryRecord({
  type: 'gotcha',
  content: 'worktree-utils dead code: shouldPruneWorktree() never called',
  area: 'cleanup',
});

MCP Server Mode (Dual CLI + MCP)

CodeGraphContext supports two operation modes:

CLI mode (batch indexing, one-shot queries):

code-graph-context index --root . --lang typescript,javascript
code-graph-context query --cypher "MATCH (f:Function) RETURN f.name LIMIT 10"

MCP server mode (live, incremental — recommended for agent use):

npx @codetiger/code-graph-context-mcp --watch --db .claude/context/data/code-graph.kuzu

With --watch, the server monitors file changes and re-indexes incrementally. No manual re-index after refactors.

Neo4j Enterprise Option

For large codebases (>100K nodes), replace KuzuDB with Neo4j:

code-graph-context index --root . --backend neo4j --uri bolt://localhost:7687

Add to settings.json:

"CodeGraphContext": {
  "command": "code-graph-context-mcp",
  "args": ["--backend", "neo4j", "--uri", "bolt://localhost:7687"]
}

Interactive HTML Visualization

Generate a browsable call graph:

code-graph-context visualize --output .claude/context/tmp/call-graph.html

Pre-Indexed Bundles

For CI/CD pipelines, ship a pre-built graph bundle with the repo:

# In CI: build + archive
code-graph-context index --root . --export .claude/context/data/code-graph.bundle.gz

# In agent startup: restore
code-graph-context restore --bundle .claude/context/data/code-graph.bundle.gz

Complexity Analysis

// Find high-complexity functions (cyclomatic > 10)
mcp__CodeGraphContext__query_graph({
  cypher:
    'MATCH (f:Function) WHERE f.complexity > $threshold RETURN f.name, f.file, f.complexity ORDER BY f.complexity DESC',
  params: { threshold: 10 },
});

Anti-Patterns

• Do not use query_graph raw Cypher before checking if a purpose-built tool covers the need
• Do not treat find_dead_code results as certain — dynamic dispatch and reflection create false positives
• With --watch MCP mode, manual re-index is not needed — avoid running index --incremental during active MCP sessions
• Graph queries do not cross language boundaries (TypeScript ≠ Python in same repo)
• Do not use KuzuDB for >100K node graphs in production — switch to Neo4j backend

When to Invoke

Skill({ skill: 'code-graph-context' });

Invoke for: impact analysis before refactoring, call chain debugging, dead code audits, module dependency reviews, and architectural dependency mapping.