An AI agent should receive a user’s tool access only after the user grants it. Vercel’s new TanStack AI integration adds that consent step to the application flow.
Vercel Connect now supports TanStack AI clients calling OAuth-protected Model Context Protocol (MCP) servers. Announced September 24, 2026, the integration provides a @vercel/connect/tanstack-ai package entry point and an connectMCPTransport adapter. Vercel says tokens are issued at runtime, so the application does not need to store user credentials in its environment.
The request path has a consent gate.
A user sends a request to the application. The server route identifies that user and prepares the MCP connection. If the user has not granted access to the target service, Connect raises a consent challenge. The route should catch that challenge and send the user through authorization before the model runs. Once consent exists, the agent can make MCP requests with a token issued for that user.
That ordering is the feature’s most useful design detail. If an application lets the model start and returns the missing-grant problem as a tool error, the user gets a confusing answer after the agent has already tried to act. Handling consent at the route boundary gives the application a clear place to stop, explain, and resume.
What the adapter handles
The adapter connects TanStack AI’s MCP client to Vercel Connect’s authorization flow. Vercel’s documentation describes requesting a fresh user-scoped token before each MCP request rather than embedding a long-lived credential in the model prompt or application environment. The developer still chooses which server to call, how to map an authenticated user to a connection, and what the product does when consent is missing or revoked.
Runtime token issuance reduces one credential-handling burden. It does not make every MCP tool appropriate for every user. Tool descriptions can be misleading, and a token’s scopes may be broader than the one action the user asked for. Review the server, scopes, tool list, and write actions before enabling them in a production agent.
A permission test worth running
- Start with a test user who has never authorized the MCP service.
- Confirm the route presents consent before model execution.
- Grant the narrowest available scope and repeat the request.
- Inspect the identity and scope used for the tool call.
- Revoke the grant and confirm the next call requires consent again.
- Attempt a write operation outside the user’s intended task and verify the application blocks it.
This test catches the difference between having a valid token and having authority for a specific action. It also gives the team a reproducible check for future changes to the MCP server’s tools.
How it fits with other Connect integrations
MustHave.ai previously covered Vercel Connect for Microsoft Teams, where revocable user access is also central. This TanStack AI release is a framework integration: it gives builders a documented way to place the same consent boundary around MCP requests in a TanStack agent. Our Cloudflare private MCP guide describes a different boundary: the network path to a server that should not be exposed publicly.
What to check before production
Verify your Vercel Connect plan, supported MCP server and framework versions against the current documentation. Keep a user-visible record of connected services and a revocation path. Log which authenticated subject approved access, when it happened, and which tool was called, without storing the token in an agent transcript. A test that covers consent, revocation, and a failed write is more informative than a happy-path demo.
Primary sources
Checked September 24, 2026. Vercel documents the consent-flow behavior; the permission test is in MustHave.ai implementation guidance.
