Skip to main content

Architecture

IronShift is built as a modular system with four interconnected components, powered by a graph-based AI agent engine.


System Overview

ComponentTechnologyRole
IDE ExtensionTypeScriptUser interface, chat panel, tool execution on user's machine, terminal integration
Backend APIPython / FastAPIAI agent orchestration, LLM routing, conversation management, tool registry
Web DashboardReact / ViteUser auth, billing, subscription management, usage analytics
Website BackendNode.js / ExpressUser management, OAuth, payment processing, API keys

Agent System

The core AI engine uses LanGraph (LangChain) to implement a stateful, graph-based agent:

ComponentPurpose
State MachineDefines agent phases, conversation state, and tool call tracking
Agent GraphMain graph with nodes for routing, LLM calls, tool execution, and approval
Cloud RouterSelects appropriate cloud provider and loads relevant tools
Tool ExecutorExecutes approved tool calls and formats results
Human ApprovalManages approval queue and user confirmation flow
Context MemoryMaintains conversation context and summary across messages
Command ValidatorValidates CLI commands before presenting to user

LLM Integration

ProviderModelsConfig Key
OpenAIGPT-5.1, GPT-5.2, GPT-5.2 Pro, GPT-5 NanoOPENAI_API_KEY
AnthropicClaude Sonnet 4.5/4.6, Opus 4.5/4.6ANTHROPIC_API_KEY
GoogleGemini 2.5 Pro/Flash, Gemini 3 Pro/FlashGOOGLE_API_KEY
xAIGrok 4XAI_API_KEY

Observability

IronShift integrates with Langfuse for LLM observability:

  • Trace every LLM call
  • Track token usage and costs per model
  • Monitor latency and errors
  • Debug agent behavior with step-by-step traces

API Endpoints

EndpointMethodDescription
/GETRoot health check — returns app status and version
/healthGETHealth check endpoint
/api/chatPOSTMain chat endpoint with streaming support