Air Pipe已验证已精选

Air Pipe turns one declarative config into a running backend — HTTP routes, cron jobs, webhooks, realtime channels and MCP tools, all from the same interface and behind the same auth.

This server hands that platform to your AI client. Claude Code, Claude Desktop or Cursor can plan a service, write the config, validate it, deploy it to staging or production, and roll it back — end to end, without leaving the chat.

The authoring loop

The interesting tool is plan_config, and it deliberately does not generate YAML server-side. It returns config-authoring instructions plus real action examples retrieved for your request, and your model writes the config. That keeps the output in your model's context where you can reason about it, rather than handing back a black box.

1. plan_config — instructions + retrieved examples for what you asked for
2. your model writes the YAML
3. validate_config — schema check; feed the error back and retry until it passes
4. create_config — persist it, optionally deploying in the same call
5. deploy_config_revision — promote or roll back to any past revision

Step 3 is a repair loop, not a gate. validate_config returns the specific schema error so the model can fix and re-submit, which is why generated configs converge instead of failing opaquely.

Tools

Discovery

Tool	Does
list_organizations	Your orgs and their organization_uuid. Call this first — the config tools need it.
list_configs	The org's configurations.
get_config	One configuration, optionally with its revision history.

Authoring

Tool	Does
plan_config	Authoring instructions + retrieved examples so your model can write the config.
validate_config	Schema-validate YAML. Returns the exact error to repair against.
create_config	Persist a config, optionally deploying it.
update_config	Update by uuid, adding a revision, optionally redeploying.
edit_config	Apply a natural-language edit to an existing config.

Lifecycle

Tool	Does
deploy_config_revision	Deploy any past revision to staging or production — this is your rollback.
undeploy_config	Remove a config from an environment.
toggle_config	Enable or disable configs.
delete_config	Delete configs.

Reusable modules — typed config fragments referenced by handle, so shared pieces (a database block, a common action) live in one place: list_modules, get_module, create_module, update_module, deploy_module_revision, undeploy_module, delete_module.

19 tools in total.

Setup

1. Create an account at https://app.airpipe.io/register — the free tier is 3,000 managed request units a month, no card.
2. Mint an MCP token from your account. It is a long-lived JWT (90 days by default), shown once, and revocable at any time.
3. Add the server:

{
  "mcpServers": {
    "airpipe": {
      "type": "http",
      "url": "https://api.airpipe.io/mcp",
      "headers": { "Authorization": "Bearer YOUR_AIRPIPE_MCP_TOKEN" }
    }
  }
}

Ask your client to list_organizations to confirm the connection.

Security

Every tool validates the token, then checks org membership, then role — three layers, and they fail closed. An agent reaches exactly what you could reach yourself with that token, and nothing more.

Tool discovery is open (registries need it), but every tool call is authenticated. Tokens are revocable by id without waiting for expiry.

What you get on the other end

A config deployed through these tools is a real service: OpenAPI docs, Prometheus metrics and OpenTelemetry traces generated from the same file, over Postgres, MySQL, SQL Server, SQLite or MongoDB. Run it on Air Pipe's managed cloud, or self-host the identical config as a single binary — the self-hosted binary serves this same MCP interface.

• Site: https://airpipe.io/
• Docs: https://docs.airpipe.io/
• Marketplace packs: https://docs.airpipe.io/marketplace