Skip to content

Connect: Vercel AI SDK

Recipe · through the TypeScript client or our MCP server · their docs checked 2026-10-30

AI SDK 7 (TypeScript). The TS client route keeps approval in Python and runs confirm tools end to end. The MCP route is one call, for tools that need no approval. There is no AI SDK for Python.

Our side

def public_app(tools: Registry, host: str, agent_token: str, reviewer_token: str) -> Any:
    """REST at /, MCP at /mcp, one bearer check for both: for a host on the internet."""
    from mcp.server.transport_security import TransportSecuritySettings

    from agent_adapter.server import create_app

    tokens = {
        agent_token: Principal("hosted-agent"),  # the token the remote framework sends
        reviewer_token: Principal("reviewer@example.com", scopes=frozenset({APPROVER_SCOPE})),
    }
    security = TransportSecuritySettings(allowed_hosts=[host], allowed_origins=[f"https://{host}"])
    return create_app(
        tools,
        authenticate=bearer_auth(tokens),
        host="0.0.0.0",  # a public bind: the public-host checks apply
        allowed_hosts=[host],  # REST: any other Host header gets 421
        mcp_options={"transport_security": security},  # MCP: the same rule
    )

import * as ai from "ai";
import { ToolClient } from "@agent-adapter/client";

const client = new ToolClient({ baseUrl: "https://tools.example.com", token: process.env.AA_AGENT_TOKEN });
const tools = await client.asVercelTools(ai.tool, ai.jsonSchema, { runId });

const result = await ai.generateText({ model, tools, prompt: "File my 2026 return." });

asVercelTools builds each tool with description, inputSchema: jsonSchema(...) and execute. A confirm tool's execute returns { pending_approval: approval_id }. Your reviewer screen then calls client.decide(approval_id, { approved }), which runs the call and returns its result. Watch the run with client.events(runId).

Their side: MCP

import { createMCPClient } from "@ai-sdk/mcp";

const mcp = await createMCPClient({
  transport: {
    type: "http",
    url: "https://tools.example.com/mcp",
    headers: { Authorization: `Bearer ${process.env.AA_AGENT_TOKEN}` },
  },
});
try {
  const { send_invoice, ...tools } = await mcp.tools(); // confirm tools stay out
  await ai.generateText({ model, tools, prompt: "Shipping rate to the US?" });
} finally {
  await mcp.close();
}
Key Meaning
transport.type: "http" Streamable HTTP
transport.url, transport.headers our /mcp endpoint and the bearer token
transport.authProvider OAuth; our bearer_auth takes static tokens, so use headers
toolApproval on generateText / streamText AI SDK 7's approval policy (needsApproval is deprecated)

Checklist

Check Why
toolApproval is a UI gate it adds a step; our executor still decides confirm tools
confirm tools go through the TS client over MCP the AI SDK does not send our resume keys back
the token stays on the server (route handler, server action) the server sends no CORS headers, and a browser bundle leaks it
MCP tool annotations are hints the AI SDK does not turn them into a policy; neither do we
everything on the public-host checklist

Sources