PROJECT
OVERVIEW
a8n is a self-hosted workflow automation platform that lets both humans and AI agents build, execute, and monitor multi-step pipelines. The architecture is a layered monolith on Next.js 16 — tRPC provides type-safe web APIs, Inngest provides durable event-driven execution with per-step retry, and a parallel MCP server exposes 22 tools for AI-client access over Streamable HTTP.
Workflow automation platform combining a visual React Flow DAG editor with durable Inngest execution and a parallel MCP server — built so humans design workflows visually while AI agents automate them programmatically.
“The same platform serves humans through a visual canvas and AI agents through the Model Context Protocol — two interaction paradigms, one execution engine underneath.”
SYSTEM
ARCHITECTURE
Engineered for production durability, type safety, and scalable domain isolation. Every module operates with strict boundaries and predictable failure handling.
MODULAR DOMAIN ISOLATION — BUILT FOR ZERO-DOWNTIME DEPLOYMENTS, TYPE-SAFE CONTRACTS, AND FAIL-SAFE EXCEPTION BOUNDARIES.
PROJECT
IMPACT
Production performance measured under live traffic load, latency stress profiles, and automated system profiling.
ALL OUTCOMES ARE EMPIRICALLY VERIFIED IN PRODUCTION ENVIRONMENTS — MEASURED UNDER HIGH CONCURRENCY, PEAK LATENCY PROFILES, AND AUTOMATED SYSTEM HEALTH AUDITS.
MCP Tools
Custom Node Types
Realtime Channels
Architecture Decision Records
KEY
CAPABILITIES
Core capabilities engineered for high-scale reliability, intuitive operator workflows, and real-time production execution.
PRODUCTION-TESTED SYSTEM FEATURES — ARCHITECTED FOR LOW-LATENCY INTERACTION, DETERMINISTIC STATE HANDLING, AND COMPREHENSIVE TELEMETRY.
Visual DAG Editor with Topological Execution
React Flow canvas with 10 custom node types (triggers, HTTP, AI providers, integrations). Before execution, the DAG is topologically sorted using the toposort library — upstream nodes always complete before their downstream consumers. Cycles are detected and rejected.
Inngest Durable Execution with Realtime Streaming
Each workflow runs as an Inngest step function with 5 durable steps. If step 3 fails, steps 1–2 don't re-execute. 9 dedicated realtime channels publish per-node execution status (loading → success → error) to the browser via SSE. Context propagates through the chain — each node's output becomes the next node's input via Handlebars templates.
MCP Server (22 Tools, 4 Resources, 3 Prompts)
Production-grade Model Context Protocol server at /api/mcp exposing workflow automation to AI clients (Claude Desktop, Cursor, MCP Inspector). Runs parallel to tRPC — tools call Prisma and Inngest directly. Bearer token auth with scoped API keys. Streamable HTTP transport.
Multi-Provider AI Nodes with Credential Scoping
AI executor nodes support OpenAI, Anthropic, and Gemini through the Vercel AI SDK. API keys are encrypted with AES-256 at rest and decrypted only at execution time. Credential queries are scoped to the workflow owner — users can only use their own keys.
5-Layer Security Architecture
Layer 1: Better Auth session management with OAuth (GitHub/Google). Layer 2: tRPC middleware (protected + premium tiers via Polar.sh). Layer 3: Row-level data isolation (every query filtered by userId). Layer 4: AES-256 credential encryption. Layer 5: Zod input validation on every tRPC procedure.
DEVELOPMENT
PROCESS
A rigorous, phased engineering lifecycle designed to transform architectural requirements into scalable, production-ready systems.
DISCIPLINED SYSTEM METHODOLOGY — ARCHITECTED FOR END-TO-END OBSERVABILITY, DETERMINISTIC REVIEWS, AND ZERO-REGRESSION RELEASE CYCLES.
TECHNOLOGY
STACK
An engineered architectural map outlining foundational nodes, runtime environments, and type-safe deployment frameworks.
DETERMINISTIC ARCHITECTURE STACK — PROFILED FOR PRODUCTION LATENCY, STRICT TYPE SAFETY, AND HIGH-CONCURRENCY SCALABILITY.