@aparte/runner-ag-ui

Render an agent served over AG-UI in aparté. The agent runs on its own server, with its own tools: Agno, LangGraph, Mastra, CrewAI, Pydantic AI, CopilotKit's runtime, or TanStack AI, whose chat route streams AG-UI. aparté draws its text, its reasoning, its tool calls and its approvals, exactly as it draws a loop it runs itself.

npm install @aparte/runner-ag-ui @aparte/core

import { AparteClient } from '@aparte/core';
import { createAgUiRunner } from '@aparte/runner-ag-ui';

const client = new AparteClient({
    streamRunner: createAgUiRunner({ endpoint: '/agui' }),
});
client.start();

• Tools that need the page (an ask_user you register with registerTool) go to the agent as AG-UI frontend tools. When it calls one, the page runs it and the run resumes with the result.
• Approvals. When the agent pauses on a call that needs a person (an AG-UI interrupt, such as TanStack AI's needsApproval), the page asks through its usual approval panel, and the answer resumes the run on the agent's server.
• Its state (STATE_SNAPSHOT, STATE_DELTA) is kept for the thread and sent back as the next run's state. CUSTOM, activities and sub-agents are checked, not drawn, in this version.
• AG-UI 1.0, held to its own corpus. The runner declares protocolVersion: "1.0", rejects the streams 1.0 tells a client to reject, drops what 1.0 does not describe (reported to onWarning), and passes the protocol's 68 conformance streams on every build.
• TanStack AI frames its runs its own way — a RUN_FINISHED per model step — so its routes take createAgUiRunner({ endpoint: '/api/chat', relaxedRunFraming: true }).

No AG-UI SDK is pulled in: the wire is plain JSON over Server-Sent Events, parsed here, with @aparte/core as the only peer dependency.

Part of the aparté monorepo. ESM-only. See Providers → AG-UI agents and TanStack AI in the docs.