This post is mainly about it's remote MCP server, because that's how I and my friends mostly use it now (via Claude or Codex)
You can add https://azurade.com/mcp in Claude or Codex (or claude code), sign in once through OAuth and the LLM will be able to generate / get all info about models and account / upload files and all that.
I've ran it through oblivious evaluator agent for 8 times to reach good agent ergonomics. After all, if I want to build an agent-native service the users (agent) should evaluate and correct it. Please feel free to point your agents at it and send me the nitpicks and gotchas they encounter. Humans are also welcome :)
Demo video (1m40s, cute dog included): https://youtu.be/tX29WLffXCk
P.S. the stack is Go and Pocketbase, official GO MCP SDK, and the frontend is statically rendered NextJS. The access token is an API key named after the client, so revoking an agent means deleting that key on your profile page. It also works like REST API with an OpenAPI spec.