Langflow

React 19 SPA using @xyflow/react canvas that enables users to compose AI workflows by dragging component nodes and connecting them via edges. The frontend maintains a real-time graph state synchronized with the backend, performing connection validation before execution to ensure type compatibility between component inputs and outputs. Changes are persisted to the database and reflected in the flow execution engine without requiring code editing.

Backend component system that discovers, catalogs, and serves component definitions (LangChain chains, custom Python classes, tool wrappers) through a registry API. Components are introspected at runtime to extract input/output types, default values, and field constraints, then serialized as JSON schemas that the frontend uses to render dynamic node UIs. New components can be added without restarting the server via the component loading mechanism.

Feature that enables voice interaction with flows by integrating speech-to-text (STT) and text-to-speech (TTS) services. User speech is transcribed to text, passed through the flow, and the output is converted back to speech. Supports multiple STT/TTS providers (OpenAI Whisper, Google Cloud Speech, etc.) and can be configured per flow. Voice sessions maintain context across multiple turns for natural conversation.

Backend data layer using SQLAlchemy ORM that persists flows, components, versions, execution history, and user data to a relational database. Supports multiple database backends (SQLite for development, PostgreSQL for production) through a unified abstraction layer. Migrations are managed via Alembic, and the schema is versioned to support upgrades without data loss.

User authentication system supporting multiple methods (local credentials, OAuth2, LDAP) with role-based access control (RBAC) for flows and components. Users are assigned roles (admin, editor, viewer) that determine permissions to create, edit, execute, and delete flows. API keys can be generated for programmatic access, and permissions are enforced at the API layer before flow execution.

System that exposes flows as webhook endpoints that can be triggered by external events (GitHub pushes, Slack messages, form submissions, etc.). Webhooks receive JSON payloads, map them to flow inputs, execute the flow, and optionally send results back to the webhook source. Webhook history is logged for debugging, and retry logic handles transient failures.

Integration with LangSmith (LangChain's observability platform) that automatically traces flow execution, component calls, and LLM invocations. Traces include latency, token usage, and error information, and are sent to LangSmith for visualization and analysis. Users can configure tracing per flow and view traces in the LangSmith dashboard without modifying flow code.

FastAPI backend service that executes flows as directed acyclic graphs (DAGs) by topologically sorting components, executing them in dependency order, and streaming execution events (start, progress, error, complete) back to the client via Server-Sent Events (SSE) or WebSocket. The engine maintains execution state in memory and persists results to the database, supporting both synchronous and asynchronous component execution with timeout and error handling.

Pre-built component patterns and starter templates that combine document loaders, text splitters, embedding models, and vector stores (Pinecone, Weaviate, Chroma, etc.) into retrieval-augmented generation flows. The system handles document ingestion, chunking, embedding, and storage through a unified interface, then provides a retriever component that queries the vector store and returns relevant documents for LLM context. Flows can be saved as templates for reuse across projects.

Framework for composing multiple AI agents that can call tools, pass messages to each other, and coordinate on tasks. Agents are implemented as components that wrap LLM instances with tool schemas, enabling function calling via OpenAI/Anthropic APIs. Message routing between agents is handled by the flow graph — agents can be connected to pass outputs as inputs to downstream agents, creating multi-turn conversations or hierarchical task decomposition patterns.

Python SDK (langflow.custom.component) that allows developers to write custom components as decorated Python classes with input/output type hints. The SDK provides decorators (@component, @input, @output) that register components with the registry, validate types at runtime, and generate JSON schemas automatically. Custom components are loaded into the registry and appear in the UI immediately, enabling rapid iteration on domain-specific logic without modifying core Langflow code.

Automatically exposes any flow as a REST API endpoint by wrapping the flow execution engine with FastAPI route handlers. Input parameters are extracted from flow input nodes and mapped to request body/query parameters, while flow outputs are serialized to JSON responses. The system generates OpenAPI schemas automatically, enabling API documentation and client code generation. Endpoints support both synchronous and asynchronous execution modes.

Database-backed versioning system that tracks flow changes over time, allowing users to create named versions (snapshots) of flows and deploy specific versions to production. Each version is immutable and includes the complete flow definition, component configurations, and metadata. Deployments can be rolled back to previous versions by selecting a prior snapshot, and version history is queryable via API for audit trails.

Interactive testing environment built into the UI that allows users to run flows with different input parameters, see outputs in real-time, and adjust component parameters without leaving the playground. The chat interface supports multi-turn conversations for flows with message history, and parameter changes are reflected immediately in subsequent executions. Execution history is retained for debugging and comparison.

Integration with the Model Context Protocol standard that enables Langflow to discover and invoke tools and resources exposed by MCP servers. MCP servers can be registered with Langflow, and their tools are automatically added to the component registry as callable components. This allows flows to interact with external systems (databases, file systems, APIs) through a standardized protocol without writing custom wrappers.