An Agent on Your Machine

A developer who runs an Agent with ms-tau in local mode gets the same chat as a deployed application: the rail, the expanded rail, streaming markdown, reasoning, tool calls, the queue, Stop, and the model picker. Only the source changes: ChatEngineProvider talks to the runtime on the developer's machine instead of the platform (ADR 099).

Start the Agent

In the Agent's project, with the ms-tau-sdk settings for local mode (its quick start lists them):

export TAU_LOCAL_MODE=true
export MAINSEQUENCE_AUTH_MODE=jwt
export MAINSEQUENCE_ACCESS_TOKEN="<your access token>"
export MAINSEQUENCE_REFRESH_TOKEN="<your refresh token>"
export TAU_LOCAL_PROVIDER="<provider>"
export TAU_LOCAL_MODEL="<model>"
uv run ms-tau

It listens on http://127.0.0.1:8787. It has no inbound authentication: whoever reaches it acts as you, with your platform credential. Keep it on this machine.

Forward it from the dev server

The page never calls the runtime directly. The Command Center SDK's localAgentProxy() (SDK ^0.5.7, from /vite) forwards the chat's routes from the application's own dev server:

// vite.config.ts
import { defineConfig } from "vite";
import react from "@vitejs/plugin-react";
import { localAgentProxy } from "@dev-mainsequence/command-center-sdk/vite";

export default defineConfig({ plugins: [react(), localAgentProxy()] });

It serves /__agent__ under vite serve only, to the page it serves, from this machine, and only for the chat's routes; it strips credentials and caller headers. MAINSEQUENCE_TAU_LOCAL_ORIGIN or localAgentProxy({ target }) points it at another port on this machine.

Mount the local source

import { ChatEngineProvider, ChatRail, ChatThread, createLocalAgentSource } from "@dev-mainsequence/command-center-ai";

const localAgent = createLocalAgentSource({ baseUrl: "/__agent__", displayName: "CRM assistant" });

<ChatEngineProvider source={localAgent} isVisible={visible} notify={notify} viewContext={context}>
  <ChatRail title="CRM assistant" onClose={close} onExpand={expand}>
    <ChatThread surface="overlay" />
  </ChatRail>
</ChatEngineProvider>

With a source, the provider takes no connection, auth, environmentUid, or defaultSession. createLocalAgentSource refuses an address on another origin. The application chooses the source, for example from its own bootstrap in development, and builds for the platform source: a production build has no /__agent__.

What differs from the platform

	Platform source	Local source
Sessions	The platform's AgentSessions	The runtime's sessions; the page's first one is the runtime's default conversation, and New session starts another
Reload	The platform's history	The runtime's history, once it serves it (ms-tau-sdk issue #47). Until then a reload starts over and the thread says so
Readiness	The platform's runtime decision, then the Agent's check	/ready, then the Agent's check. An Agent that is not running is unavailable, and the chat looks for it again every few seconds
Model picker	The session's model on the platform	The session's model on the runtime. A new conversation shows "Configured by the local Agent" and its first message uses the model the runtime was started with; a model chosen before then applies from the next message
Archive, search, insights, provider settings, Agent icons	Yes	Not offered; the explorer and footer hide them

useChatEngine().capabilities says which of these the mounted source offers.

Try it without an Agent

The package's standalone application runs the local source against a scripted runtime: npm run ai:dev and open /?stand-in&local. Open /?local to talk to a real ms-tau on 127.0.0.1:8787 through the dev server's localAgentProxy().