One App, two kinds of users

Say you keep a recipe box. You want to flip through it on a Sunday. You also want your Agent to add the soup it found and pull up everything with leeks.

In most setups that’s two projects: a website for you, and a separate “tool” for the AI that talks to the same database in its own way. They drift apart. One day the AI writes a recipe the website can’t show.

In Fungi it’s one App.

Two front doors, one house

Every Fungi App is built on the same contract, called Agar. An App has:

The Agent doesn’t scrape your page or guess at buttons. It calls the method, gets typed data back, and the host checks its permission on every call.

Here’s what that looks like

export const appBackend = defineAppBackend({
  methods: {
    search: defineAppMethod({
      description: "Find recipes that use an ingredient.",
      effect: "read",
      requires: [],
      input: z.strictObject({ ingredient: z.string().min(1) }),
      output: z.strictObject({ titles: z.array(z.string()) }),
      handler: async ({ ingredient }) => ({ titles: await findBy(ingredient) }),
    }),
  },
});

The build reads that definition and writes the App’s API for you. On Oyster, Agents find it through one MCP endpoint and call it as env.RECIPES.search({ ingredient: "leeks" }). Claude Code, Codex or any MCP client can do the same with a key you give it.

Why it matters

Honest limits

Build Apps for everyone · Agar docs