MCP server

Start the server

Build the Software Graph first, then start the MCP server. It runs as a stdin/stdout process using a line-JSON protocol. The server exposes capabilities — structured tools that agents can invoke.

pnpm exec ontoly build .
pnpm exec ontoly mcp
pnpm exec ontoly mcp --list  # list available capabilities

Graph capabilities

Core capabilities for navigating and inspecting the graph structure.

• FindNode, FindFunction — locate specific nodes by name or ID
• FindDependencies, FindDependents — dependency traversal
• TraceExecution — follow call chains
• InspectFile, InspectModule, InspectClass, InspectFunction — detailed inspection
• FindCycles — detect circular dependencies
• FindDeadCode — find unreachable code
• FindEntrypoints — locate application entry points
• FindConfiguration — find configuration nodes
• GraphStatistics — graph metrics and counts

Intelligence capabilities

Higher-level capabilities that combine graph traversal with semantic analysis.

• ExplainArchitecture — architecture summary
• ImpactAnalysis — blast radius computation
• TraceRequestLifecycle — end-to-end request tracing
• FindFeatureOwner — feature ownership analysis
• FindAuthenticationFlow — authentication pattern detection
• FindResponsibleFunction — responsibility attribution
• FindDatabaseAccess — database access patterns
• FindConfigurationUsage — configuration dependency analysis
• EvidencePack — bounded evidence for agent context
• ImplementationPlan — graph-backed refactoring plan

Search capabilities

Capabilities for semantic search, intent resolution, and concept discovery.

• SearchConcept — concept-level search against the Semantic Index
• FindFeature — feature-level touchpoint discovery
• ResolveIntent — parse natural language into a query plan
• SemanticNeighborhood — explore conceptual neighborhoods
• FeatureOwnership — ownership analysis for a feature
• IntentExpansion — expand a query into related concepts

History capabilities

Capabilities that combine graph structure with Git history.

• History — full history intelligence for a node or query
• Ownership — likely code owners based on history and structure
• Hotspots — high-churn, frequently-modified nodes
• Cochanges — files that commonly change together
• Stability — hotspot, churn, and stability scores

Agent workflow pattern

Agents should follow a disciplined query sequence: resolve → locate → inspect → impact → evidence → synthesize. Never skip straight to evidence. Start narrow and widen only when needed.

• Search for the concept to get candidate nodes
• Select a stable node from the candidates
• Inspect its immediate neighborhood
• Run scoped impact analysis
• Request a bounded evidence pack
• Fall back to source only when evidence is incomplete