AutomationCAN I VIBECODE
Relay.app?
BUILD FIRST
- One specific customer type and one primary use case
- A vertical slice that completes the main job without manual workarounds
- Visible limits, review steps, and a fallback when an external dependency fails
LEAVE FOR LATER
- Broad third-party integrations
- High-scale collaboration, syncing, and permissions
- Automation that can cause harm without human review
RESEARCH NOTES
What this tool actually does.
Relay.app is a visual automation platform combining no-code workflow building with AI capabilities. Users set up triggers from connected apps and build multi-step workflows using Paths—conditional branches with rules that route data, merge back together, and support test runs with real data. AI steps come in three forms: built-in actions (extraction, summarization, translation, transcription), custom prompt steps for text or structured output, and agentic tools where AI decides which actions to run from an allowed set. Models can be used via Relay.app's own AI credits (OpenAI, Anthropic, Google) or by connecting provider API keys directly. Human-in-the-loop review lets users approve AI step outputs before the workflow proceeds. The platform is designed for less technical users, avoiding raw JSON/API/LLM complexity, and includes collaboration features and an assistant to help debug prompt
RESEARCHED BUILD PROMPT
Give this to your coding agent.
Build a web app called 'FlowRelay' (internal codename), a no-code automation platform inspired by the workflow of Relay.app: trigger → conditional paths → actions → merge → AI steps → human approval. The target user is a non-technical business owner; no JSON, API keys, or LLM configuration should ever appear in the UI.
Core workflow: User selects a trigger (webhook, form submit, email, schedule), then builds a workflow on a visual canvas. Key steps: (1) basic actions (send email, update CRM, create task), (2) AI steps with three modes: built-in (summarize/extract/translate/transcribe), custom prompt with structured output, and agentic (AI picks from a list of allowed actions), (3) Paths: conditional branches with rules over variables/previous outputs, mergeable back into one flow. Human-in-the-loop: for any AI step, the user can toggle approval mode; workflow pauses and shows a review queue item.
User-facing requirements: drag-and-drop canvas, step configuration panels with plain-language labels, a test runner that executes the workflow against sample data and shows branch paths taken, a history log, and an assistant panel that suggests prompt improvements.
Technical requirements: Node.js + TypeScript backend, React + React Flow for the editor, PostgreSQL for workflow definitions/execution state, a worker queue (BullMQ) for running workflows, an OpenAI-compatible AI adapter layer (support provider-agnostic calls), and a static webhook receiver for trigger testing. Store workflow JSON with versioning. Implement approval as blocked runs table.
Implementation approach: use React Flow with custom nodes for paths/merge; model the workflow as a DAG executed by a rule-based engine rather than a general workflow engine, to keep conditional logic transparent. Provide seeded demo workflows.
Do NOT copy Relay.app's brand, logo, product name, marketing copy, or visual design. Use an original color palette and layout. Out of scope: third-party app integrations beyond a mock CRM/email, multi-tenant accounts, SSO, mobile apps, public marketplace, audit logs, and scheduling beyond a simple interval.
Include a README with setup steps (database migration, env vars for AI provider keys), architecture overview, and important limitations (no production security, mock integrations only, AI output nondeterminism, no partial reruns, single-tenant).