Harnesses

A harness is a hosted runtime for an AI SDK endpoint at its own origin. Conversations stay on the cloud, and the hosted proxy calls your endpoint for each run.

The Harnesses page lists each harness's origin, backend, thread count, active count, and last update. Open a harness to connect an app, configure which endpoints the proxy may call, and observe its live instances and conversation ledger.

How it works

The harness origin is built from the harness id and the configured harness domain. A browser connects to that origin with a credential and a cloud workspace and thread identity. For every run, the proxy chooses an allowed backend endpoint and calls your AI SDK route while the conversation remains in Assistant Cloud.

Harnesses need two forms of availability. The hosted harness service must be enabled for the dashboard, and the project plan must include the harnesses feature. When the service is unavailable, Harnesses is absent from the sidebar, the pages are unavailable, and reads or writes answer Hosted harnesses are not enabled. When a project lacks the plan feature, the page shows its plan required state and the project cannot read or change harnesses.

Configure

Create a harness by entering a trimmed name from 1 to 80 characters. The cloud creates an AI SDK harness with a generated id of 21 lowercase letters and digits, an empty endpoint allowlist, and Allow localhost on.

Endpoint entries are trimmed, empty entries are discarded, and duplicates collapse to one entry. An entry is an absolute HTTP or HTTPS origin with an optional path prefix and an optional leading *. hostname wildcard. A public, nonloopback endpoint must use HTTPS. *.example.com matches subdomains, and a path prefix limits the part of the endpoint that may be called.

Loopback hosts, including localhost, any 127.x.x.x address, and ::1, belong only behind Allow localhost. Realtime Voice Provider lists Off and the project's OpenAI providers, and saves its choice immediately.

What the page shows

The detail page has Connect, Observe, and Configure tabs. Its Facts rail shows the copyable Origin and Harness ID, Type as AI SDK, Voice as the provider name or Off, Credentials, Anonymous access, Created, and Updated.

The Connect tab offers an assistant-ui snippet and an Other snippet; the tab fills in the harness origin and the endpoint path. In both, url is the path of the first allowed endpoint, or /api/chat when the allowlist is empty. origin is the harness origin. credential is a project API key or a token minted with POST /v1/auth/tokens. workspaceId identifies the workspace whose conversations are opened. The Other snippet also takes threadId, the conversation to open; the assistant-ui snippet lists the workspace's threads instead.

import { AuiConfig, AuiProvider } from "@assistant-ui/react";
import { HarnessThread, HarnessCloudThreadList } from "@assistant-ui/react-harness-sdk";

const config = AuiConfig({
  threads: HarnessCloudThreadList({
    url: "/api/chat",
    origin: "https://harness.example.com",
    credential: () => fetchCredential(),
    workspaceId: "ws_...",
  }),
});

export function Chat() {
  return (
    <AuiProvider config={config}>
      <HarnessThread />
    </AuiProvider>
  );
}

import { useHarness, HarnessCloud } from "harness-sdk";

const { messages, sendMessage } = useHarness({
  transport: HarnessCloud({
    url: "/api/chat",
    origin: "https://harness.example.com",
    credential: () => fetchCredential(),
    workspaceId: "ws_...",
    threadId,
  }),
});

Neither harness package is published during the preview, so these snippets show the API the Connect tab hands out, not an install you can run yet.

Costs and limits

Harnesses are available on the following plans. Plan limits can also be overridden for a project.

When a project's plan loses the feature, the cloud suspends every harness in that project. When the plan regains the feature, it clears that suspension and restores the harnesses. Creation rechecks the plan and cap inside its transaction. At the cap, the dashboard refuses creation with This plan includes {n} harness or This plan includes {n} harnesses.

Observe runs and conversations

Live instances shows the instances the hosted proxy holds for the harness. A client attached to a thread or thread list makes an instance appear, and an instance stays for five minutes after its last client leaves. The page polls every 5 seconds, waits 30 seconds after an error, and does not retry automatically.

The live status shows the total connections and pulses while that total is above 0. If no client is attached, the empty state explains that no instance is active. If the registry is unavailable, the page shows The registry is unreachable and a Retry action. If the registry is not configured, the Live section is not shown.

The threads ledger lists the 50 most recently updated harness thread rows and shows {active} active, {total} total.

Deleting a harness

Choose Delete to stop threads on the harness and remove their cloud history. Deletion removes the harness thread records as well. The audit log records harness.create, harness.update, and harness.delete; the delete entry includes the harness name, endpoints, localhost setting, and voice provider.

Troubleshooting