MCP Server
Model Context Protocol integration that gives AI assistants direct access to your knowledge graph. Claude Desktop, Claude Code, and other MCP-compatible clients can search, explore, and manipulate your knowledge.
Setup
# Generate MCP credentials
kg oauth create-mcp
# Add to Claude Desktop config (output includes ready-to-paste config)
What AI Agents Can Do
Search Your Knowledge
"Find concepts related to distributed systems"
"What does the knowledge graph say about eventual consistency?"
"Search for evidence about CAP theorem tradeoffs"
The agent uses semantic search to find relevant concepts, then retrieves evidence quotes from source documents.
Explore Relationships
"How does 'microservices' connect to 'deployment complexity'?"
"What concepts are related to 'event sourcing'?"
"Find the path from 'technical debt' to 'team productivity'"
The agent traverses your knowledge graph to discover connections you might not have considered.
Validate Claims
"What evidence supports the claim that caching improves performance?"
"Are there any contradicting viewpoints about serverless architecture?"
"Show me the sources for concepts about API design"
Every concept has grounding strength and evidence quotes. The agent can assess how well-supported a claim is.
Build Knowledge
"Create a concept for 'blue-green deployment' in the devops ontology"
"Add a relationship: 'feature flags' ENABLES 'gradual rollout'"
"Ingest this text about Kubernetes patterns"
The agent can create concepts, add relationships, and submit documents for extraction.
Manage Jobs
"What ingestion jobs are running?"
"Approve the pending job for the architecture document"
"Cancel the failed job"
Full control over the extraction pipeline.
Available Tools
Query Tools
| Tool | What It Does |
|---|---|
| search | Find concepts, sources, or documents by semantic similarity |
| concept (details) | Get full concept info with all evidence and relationships |
| concept (related) | Find concepts connected within N hops |
| concept (connect) | Discover paths between two concepts |
Graph Tools
| Tool | What It Does |
|---|---|
| graph (create) | Create concepts or relationships |
| graph (edit) | Update existing concepts or edges |
| graph (delete) | Remove concepts or relationships |
| graph (list) | Query concepts/edges with filters |
Ingestion Tools
| Tool | What It Does |
|---|---|
| ingest | Submit text, files, or directories for extraction |
Management Tools
| Tool | What It Does |
|---|---|
| ontology | List, inspect, or delete knowledge domains |
| job | Monitor, approve, cancel, or delete jobs |
| document | List documents, retrieve content |
| source | Access original source text or images |
| artifact | Retrieve saved analysis results |
Analysis Tools
| Tool | What It Does |
|---|---|
| analyze_polarity_axis | Project concepts onto semantic dimensions |
| epistemic_status | Check knowledge validation state |
Example Conversations
Knowledge Discovery
User: "What does our knowledge graph say about database scaling?"
Agent: Uses search tool → Finds concepts about sharding, replication, read replicas, connection pooling → Uses concept details → Retrieves evidence quotes → Synthesizes answer with citations
Research Assistance
User: "I'm writing about microservices. What related topics should I cover?"
Agent: Uses search → Finds "microservices" concept → Uses related → Discovers connected concepts: service mesh, API gateway, distributed tracing, eventual consistency → Suggests outline based on knowledge graph structure
Fact Checking
User: "Is it true that event sourcing improves auditability?"
Agent: Searches for evidence → Finds grounding strength of 0.85 (well-supported) → Retrieves evidence quotes → Shows sources confirming the claim with specific quotes
Knowledge Building
User: "Add what I just learned about CQRS to the architecture ontology"
Agent: Uses ingest tool → Submits text → Monitors job → Reports when extraction completes → Shows new concepts created
Performance Notes
- Search: ~200ms including vector similarity
- Graph traversal: ~150ms for 2-hop relationships
- Path finding: Use threshold ≥ 0.75 to avoid slow queries
- Batch operations: Up to 20 ops per queue
Lower similarity thresholds (0.60-0.74) are slower but more exploratory. Values below 0.60 can cause timeouts.