Skip to main content

Configure your app with AI

Appilot apps are configured by describing a content model: the views, controls, forms, action plans, and knowledge the agent reads to act inside your app. Authoring that by hand is easy to get subtly wrong: a flow that opens a field but never submits, a selector that breaks on the next render, half-translated labels, a knowledge article that is really a procedure.

The app-configurator turns that into a conversation. You connect your AI tool to Appilot, then ask it to set up, audit, fix, and validate your app's configuration, checked against a formal config health contract. In ChatGPT and claude.ai that connection is a URL you paste and approve once; nothing is installed. Coding agents and Claude Desktop can run the same tools locally instead, alongside the skill that carries the full procedure.

What it does​

  • Reads your app's content model and explains it in plain language.
  • Validates it against the health contract and returns a prioritized, severity-ranked report (critical → low), each finding with the exact fix.
  • Authors new configuration one entity at a time: views, controls, forms, tools, zones, action plans, knowledge articles and session templates. It asks the instance which values each field accepts rather than guessing them.
  • Looks at your screens and ranks locator candidates by whether they will still resolve after your next deploy, from the live page or from markup you paste.
  • Fixes issues with your consent. The server re-validates every write.
  • Soaks the live page to confirm selectors and plans actually resolve against the rendered DOM.
  • Records what is missing. If it hits a gap in Appilot itself, it can file a report. That report is read; a reply is part of a support plan rather than something the tool promises.

Works with your tools​

ChatGPT and claude.ai reach it over HTTPS. Neither can start a program on your machine, so both take the endpoint as a custom connector and authenticate with OAuth, and you supply your service token once, on a consent screen. The setup is four steps and no download: appilot.space/connect.

Coding agents run it locally. The skill follows the open Agent Skills standard, so it installs into Claude Code, Codex CLI, Cursor, and other standard-compatible agents: the same file, no per-tool variants. Claude Desktop installs it as a single double-clickable bundle from appilot.space/downloads.

Without the MCP server the skill still works in advisory mode: paste or export your config and it audits it by reasoning against the contract.

Cloud or on-premise​

The MCP server is endpoint-agnostic. Point it at a cloud tenant or your own on-premise / sovereign instance with one setting; the skill never hard-codes a host. The static validation runs locally with no network, so it works air-gapped. A fully sovereign setup (your instance, your model, this tooling) can run with no outbound traffic.

What the connected client receives​

Connecting a client hands it your configuration. When it reads or exports the content model it receives your views, controls, forms, tools, zones, action plans, session templates, and the full text of your knowledge base articles. That content then lives in the client's conversation, under that vendor's terms rather than ours, and neither you nor we can reach it or delete it there. Decide that before you connect a shared or consumer account.

Two things never travel over the remote connection: the minted widget secret and the credentials stored for outbound HTTP-proxy tools, which the server refuses to send. End-user conversation history is not exposed by any tool in this surface.

Access is scoped by service token, and the four scopes are config:read, config:write, provision:write and feedback:write. Issue a read-only token when all you want is an audit.

Next steps​