Build with AI
There are two ways to use AI with Air Pipe. Don't confuse them:
- Build Air Pipe with AI — describe the API you want in plain English and get a working, validated config (this page).
- Expose your API to AI — turn your own interfaces into tools an AI agent can call. See MCP tools.
1. Generate a config from a description
The Air Pipe app has a Build with AI panel. Describe what you want — "a REST
API over my Postgres orders table with JWT auth" — and it returns a complete
Air Pipe config.
Under the hood it does more than prompt an LLM:
- Retrieval over the marketplace corpus + config schema, so the output uses real Air Pipe patterns, not hallucinated fields.
- Schema validation with auto-repair — the generated YAML is validated against the engine schema and repaired before it's handed back, so what you get is deployable, not a sketch.
- Backed by frontier models (Claude / DeepSeek). Requires being signed in.
The result is a normal config — review it, tweak it, and deploy it like any other. It's a starting point you own, not a black box.
2. Manage Air Pipe from your AI client (MCP)
Air Pipe also exposes a management MCP server, so an MCP-capable client (Claude Code, Claude Desktop, Cursor, …) can author, validate, and deploy configs conversationally. Point your client at Air Pipe's MCP endpoint with a bearer token and these tools become available:
| Tool | Purpose |
|---|---|
list_organizations | List orgs you can deploy to |
list_configs / get_config | Browse existing configs |
plan_config | Generate a config from a description (retrieval + validate/repair) |
validate_config | Schema-check a config without saving |
create_config / update_config / edit_config | Author and modify configs |
deploy_config_revision / undeploy_config / toggle_config | Ship it |
A typical flow: ask your assistant to "build and deploy an API that…", it calls
plan_config to draft + validate, create_config to save, then
deploy_config_revision to go live — all without leaving the chat.
This management MCP is different from exposing your own interfaces as MCP tools. Here, the Air Pipe platform is the thing being driven by AI; there, your API is the thing an agent calls.