AI-Native Intelligence: The Three Levels
Reaktor is the first application engine where AI intelligence is a first-class primitive, not an afterthought. AI integration operates at three levels, each building on the one below.
Level 1: Reaktor MCP Server
The reaktor-mcp module is a proper Model Context Protocol (MCP) server that exposes Reaktor's internals as tools for any MCP-compatible LLM (e.g., Claude Code, Cursor).
- Graph Introspection: Query the running graph's topology, node states, and port types.
- Schema Definitions: Retrieve Schema DSL types and their validation rules.
- Module Registry: List available modules and their APIs.
- Documentation: Access full API documentation and example patterns.
- Build & Deploy: Trigger builds, run tests, and deploy specific runtimes.
Level 2: Blueprint Copilot
The Blueprint Copilot is an AI assistant integrated directly into the Blueprint editor. It uses the MCP server for context and adds editor-specific intelligence.
- Node Implementation Generation: Automatically generate a complete, compilable node implementation based on the node's ports and surrounding graph context.
- Adapter Node Generation: When a developer draws an edge between mismatched types, the Copilot generates a transformation node that transforms type B into type A.
- Schema Evolution: Identifies affected nodes and ports when a Schema DSL type is modified, generating migration code and highlighting breaking changes.
- Graph Pattern Suggestions: Recognizes common patterns (e.g., repository + interactor + screen) and suggests completing them.
Level 3: Runtime AI Nodes (reaktor-ai)
An AI node in the graph is just another node with typed ports. The graph runtime handles its lifecycle, caching, fallbacks, and telemetry identically to any other node.
| Node Type | Purpose | Use Case |
|---|---|---|
| LLMNode | Prompt + context in, structured response out. | Natural language processing and content generation. |
| EmbeddingNode | Text or document in, vector embedding out. | Semantic search and similarity matching. |
| ClassifierNode | Input data in, classification + confidence out. | Content moderation and intent detection. |
| RAGNode | Query + retrieval context in, augmented response out. | Knowledge-grounded Q&A. |
| AgentNode | Goal + available tools in, action sequence out. | Multi-step reasoning and automated workflows. |
How They Compose
The three levels form a feedback loop:
- MCP Server provides the context (the graph, schemas, and docs).
- Blueprint Copilot uses that context to generate code and manage the graph.
- Runtime AI Nodes execute at runtime and feed their telemetry back into the MCP server, allowing the Copilot to suggest optimizations.