Contexts

Agents

Agents are the newest place software runs, and the reason the Continuum was worth building now. When a model can call your application, the caller is no longer a person or a service. It is a reasoning system. Stone.js treats it like any other context: a cause, an intention, a resolution.

#An agent is a cause like any other

The principle

A tool call from an agent is structurally identical to an HTTP request: a named intention with arguments, expecting a result. If your architecture already separates domain from context, exposing it to agents is a new adapter, not a new codebase.

In Stone.js

There are two agents here: the one that calls your app at run time, and the coding agent that builds it with you. Both are contexts. The one that ships today is the builder: @stone-js/mcp-dev serves it the framework and your app; exposing the running domain as tools is the next context, coming.

#Serve the agent that builds with you

Add @McpDev() and the manifest gains one command, stone mcp. It starts an MCP server (stdio) that gives your coding agent the framework's knowledge and a live, read-only view of this app: its routes, commands, providers and config.

app/Application.tsdeclarativeimperative

import { StoneApp } from '@stone-js/core'
import { Routing } from '@stone-js/router'
import { McpDev } from '@stone-js/mcp-dev'

@McpDev()   // 'stone mcp' serves the framework + this app to your coding agent
@Routing()
@StoneApp()
export class Application {}
import { defineStoneApp } from '@stone-js/core'
import { routerBlueprint } from '@stone-js/router'
import { mcpDevBlueprint } from '@stone-js/mcp-dev'

export const App = defineStoneApp(
  { name: 'tasks' },
  [routerBlueprint, mcpDevBlueprint]
)

agent sessiontext

agent → tools/list
stone ← stone_routes · stone_app · stone_search · stone_docs
agent → stone_routes
stone ← GET /tasks (Tasks.list) · POST /tasks (Tasks.create)
agent → stone_search { query: "how do adapters collapse at runtime" }

The agent reads your app the way the framework does: one blueprint, one source of truth.

The full tool list, the --init flow that writes .mcp.json, and the bundled Agent Skills are in the MCP dev server docs. A machine-readable llms.txt map of the docs is served at the site root, so any model can load the whole mental model at once.