n8n

n8n implements a directed acyclic graph (DAG) execution model where users compose workflows by connecting nodes on a canvas. The system parses workflow definitions into an execution plan, manages data flow between nodes through a context-aware expression system, and executes nodes sequentially or in parallel based on connection topology. The Workflow Execution Engine (packages/workflow, packages/core) handles lifecycle management including initialization, execution, error recovery, and state persistence across distributed runners.

n8n provides a node registry system (Node Type System and Registration) where each integration is a self-contained node package with credential handling, API client initialization, and parameter validation. The Credential System (packages/cli, packages/core) supports dynamic credential injection via environment variables or external secret managers, with OAuth2 flows handled through @n8n/client-oauth2. Nodes are loaded at runtime from packages/nodes-base and community packages, with type definitions ensuring parameter safety.

The Project-Based Authorization and Sharing subsystem enables workflows to be organized into projects with granular access control. Users can be assigned roles (viewer, editor, owner) per project, controlling read/write/delete permissions. Workflows within a project share credentials and can reference each other. The system tracks who created/modified workflows and when. Audit logs record all user actions for compliance.

n8n provides nodes for interacting with vector stores (Pinecone, Weaviate, Milvus, Chroma) enabling retrieval-augmented generation (RAG) workflows. Nodes support document embedding via LLM providers, vector storage, and semantic search. The system handles chunking, metadata filtering, and result ranking. Integration with LLM nodes enables RAG chains where retrieved documents augment LLM prompts.

The Source Control and Environment Management subsystem enables workflows to be version-controlled via Git integration. Workflows can be exported to Git repositories, with each workflow as a separate file. The system supports branching, merging, and environment-specific configurations. Credentials and sensitive data are excluded from Git, stored separately in n8n. Deployment workflows can pull from Git and deploy to different environments.

The Node Development and Community Packages subsystem provides a TypeScript SDK (@n8n/node-dev) for building custom nodes. Developers define node metadata (name, description, properties), implement execute() method, and publish to npm. Custom nodes are loaded at runtime from npm packages, enabling community contribution. The system validates node structure, provides type definitions, and handles credential binding.

The Observability and Telemetry subsystem provides execution logs with step-by-step results, timing information, and error traces. Metrics track execution count, success rate, and duration. The system integrates with external monitoring tools (Prometheus, Datadog, New Relic) via webhooks and API. Alerts can be configured to notify on failures, slow executions, or anomalies. Execution history is queryable by workflow, status, date range.

The Chat Hub Backend System and Chat Hub Frontend provide a conversational interface for interacting with workflows. Users can chat with workflows, triggering them with natural language input and receiving results in conversation format. The system supports multi-turn conversations with context preservation. Workflows can be configured to expose chat interfaces for end-user interaction.

n8n integrates LangChain through packages/@n8n/nodes-langchain, providing nodes for LLM calls, prompt templates, memory management, and tool binding. The AI and LangChain Nodes subsystem allows users to compose chains visually by connecting LLM nodes to retrieval nodes, memory nodes, and tool nodes. Expression system supports dynamic prompt injection, and execution engine manages token counting and streaming responses. Supports OpenAI, Anthropic, Ollama, and other LLM providers via credential system.

The AI Workflow Builder (AI Features subsystem) accepts natural language descriptions and generates workflow definitions using an LLM. The system parses user intent, maps to available nodes and integrations, generates node configurations, and creates connections. Generated workflows are editable in the canvas before execution. This leverages the expression system and node registry to ensure generated workflows are syntactically valid and executable.

The Instance AI system (AI Agent and Tool Execution, Instance AI — Autonomous Agent System) enables workflows to include autonomous agent nodes that use planning and reasoning to accomplish goals. Agents receive tool definitions from connected nodes, generate execution plans using an LLM, call tools iteratively, and adapt based on results. The system manages agent state, tool invocation, and error recovery. Agents can operate in single-step or multi-step modes with configurable max iterations.

The Workflow Data Access and Expression System (@n8n/expression-runtime, packages/workflow) provides a JavaScript-like expression language for transforming data between nodes. Expressions support variable interpolation, function calls, conditional logic, and array/object manipulation. The expression editor includes autocomplete based on upstream node outputs, type inference, and syntax validation. Execution engine evaluates expressions in a sandboxed context with access to workflow data, environment variables, and built-in functions.

The Distributed Execution and Scaling subsystem uses a task-runner abstraction (packages/@n8n/task-runner) to decouple workflow execution from the main process. Workflows can be executed on the main process, in worker threads, or on remote task runners. The system manages job queues, worker pool management, and result aggregation. Execution recovery ensures failed tasks are retried with exponential backoff. This architecture enables horizontal scaling by adding more task runners.

n8n provides webhook nodes that expose HTTP endpoints for triggering workflows. Webhooks support multiple HTTP methods, request validation via JSON schema, and response mapping to customize HTTP responses. The system generates unique webhook URLs per workflow, manages authentication (basic auth, API key, OAuth), and logs incoming requests. Webhook execution is asynchronous — requests return immediately while workflow executes in background.

n8n provides schedule trigger nodes supporting cron expressions, interval-based scheduling, and timezone-aware execution. The scheduler runs in the main process or can be delegated to task runners. Execution history tracks scheduled runs with timestamps and results. The system supports one-time schedules and recurring patterns with configurable retry behavior on failure.

The Execution Recovery and Error Handling subsystem provides error nodes that catch failures from upstream nodes and trigger alternative execution paths. Workflows support retry policies with exponential backoff, max retry counts, and conditional retry logic. Error callbacks allow workflows to notify external systems of failures. The system tracks error context (node name, error message, stack trace) and makes it available to error handlers.