Air Pipe turns AI features into endpoints your product calls — RAG, an LLM gateway and MCP tools running on your own database, authenticated and traced like any other route. The model is a config value you can swap, and you are billed per request rather than per AI credit.
Stand up a RAG chatbot on Postgres + pgvector or MongoDB Atlas Vector Search, put a governed LLM gateway in front of any model, and expose the same interfaces as MCP tools — all as config, on your infrastructure.
The model, the API key and the system prompt live in configuration. Point an endpoint at a different provider without changing its shape, and keep the contract your product already calls.
Retrieval is where RAG quietly goes wrong. Every AI endpoint ships OpenTelemetry traces, Prometheus metrics and OpenAPI docs, so you can see which step ran and how long it took rather than inferring it from a bad answer.
Gumloop meters credits that scale with the model tier and the nodes you run. Air Pipe meters one request unit per request — a whole request, whatever ran inside it. It's a different axis, and it's the one that stays predictable as the AI work gets heavier.
Managed or self-hosted. Self-host it and your documents, your embeddings and your database never leave your network.
Beyond AI endpoints, the same config serves realtime — WebSocket and MQTT pub/sub with presence, history-rewind and retained messages — that your product's users connect to directly. A no-code AI workflow builder produces triggered or batch runs, not a realtime channel layer your app subscribes to. Here it's one backend: AI endpoints, automation and realtime in the same artifact, and your pipeline runs on each message.
If the person building the automation isn't an engineer, or you want a library of ready-made scraping and enrichment nodes on a visual canvas, Gumloop fits better — we don't have either. Air Pipe suits you when the AI feature has to be a service you own: authenticated, documented, traced, on your data.