Most "add MCP to your app" posts stop at stdio: a local process, a config file, remote I shipped one for Deoochform, a form builder where the Here is the whole thing, Next.js 16 App Router, no framework beyond the official Four HTTP surface...
Most "add MCP to your app" posts stop at stdio: a local process, a config file,
a token you paste in. That works on your laptop and nowhere else. A remote
MCP server, the kind Claude or ChatGPT connects to over HTTPS with a browser
sign-in, needs OAuth, and the interesting parts are the ones no tutorial covers.
I shipped one for Deoochform, a form builder where the
whole point is that you can build the form by asking an assistant instead of
dragging fields around. So the MCP server is not a side feature, it is the
product surface. That forced me to get the auth story right rather than
hand-waving it with a pasted API key.
Here is the whole thing, Next.js 16 App Router, no framework beyond the official
SDK.
The shape of it
Four HTTP surfaces:
POST /api/mcp: the MCP endpoint itself.
GET /.well-known/oauth-protected-resource/api/mcp: "here is who authorizes me".
GET /.well-known/oauth-authorization-server: "here are my OAuth endpoints".
/authorize, /token, /register: the OAuth endpoints themselves.
A client that has never seen your server walks all four in order, unprompted.
That discovery chain is the whole reason a user can type a URL into Claude and
get a browser sign-in instead of a token prompt.
1. The MCP endpoint
The SDK ships a transport that speaks Web-standard Request/Response, which is
exactly what an App Router route handler deals in:
import { WebStandardStreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/webStandardStreamableHttp.js";
async function handle(request: Request) {
const actor = await resolveActor(request);
if (!actor) return unauthorized(request);
const server = createMcpServer(actor);
const transport = new WebStandardStreamableHTTPServerTransport({
sessionIdGenerator: undefined,
enableJsonResponse: true,
});
await server.connect(transport);
return transport.handleRequest(request);
}
export { handle as GET, handle as POST, handle as DELETE };
Two things worth pausing on.
sessionIdGenerator: undefined makes the server stateless. On Vercel or any
serverless host, consecutive requests land on different instances, so there is
nowhere for a session to live. Stateless is not a downgrade here, it is the only
thing that works.
The server is constructed per request, with the actor baked in. Not a
module-level singleton with the user passed into each tool call. Every tool
closes over the caller's identity, so there is no code path where a tool can
read a row belonging to somebody else. It costs one object allocation per
request and removes an entire category of bug.
CORS, or: why your connector silently never connects
Browser-based clients (ChatGPT, Claude on the web) call your endpoint
cross-origin. The browser fires a preflight OPTIONS first, and if that fails
the real request never happens. You see nothing in your logs, and the client
says something unhelpful about being unable to connect.
const CORS_HEADERS = {
"Access-Control-Allow-Origin": "*",
"Access-Control-Allow-Methods": "GET, POST, DELETE, OPTIONS",
"Access-Control-Allow-Headers":
"Content-Type, Authorization, mcp-session-id, mcp-protocol-version, Last-Event-ID",
"Access-Control-Expose-Headers": "mcp-session-id, mcp-protocol-version",
};
Expose-Headers is the one people miss. Without it the browser hides the MCP
protocol headers from the client even though your server sent them.
2. The 401 that starts the dance
An unauthenticated call must not just return 401. It has to say where to go,
using WWW-Authenticate per RFC 9728:
const metadataUrl =
`${origin}/.well-known/oauth-protected-resource/api/mcp`;
return Response.json(
{ error: "Unauthorized" },
{
status: 401,
headers: { "WWW-Authenticate": `Bearer resource_metadata="${metadataUrl}"` },
},
);
This single header is the trigger for the entire browser sign-in flow. Return a
bare 401 and the client concludes your server is broken rather than
password-protected.
Note the metadata path: RFC 9728 nests it under the resource's own path. If you
serve two MCP endpoints, /api/mcp and /api/mcp/v2, each needs its own
document at its own nested path. They are not interchangeable.
// /.well-known/oauth-protected-resource/api/mcp/route.ts
export async function GET() {
return NextResponse.json({
resource: `${origin}/api/mcp`,
authorization_servers: [origin],
});
}
3. Being your own authorization server
You are probably already sitting on a session system. You do not need Auth0 for
this. The metadata document is a static JSON file plus three routes:
export async function GET() {
return NextResponse.json({
issuer: origin,
authorization_endpoint: `${origin}/authorize`,
token_endpoint: `${origin}/token`,
registration_endpoint: `${origin}/register`,
response_types_supported: ["code"],
grant_types_supported: ["authorization_code"],
code_challenge_methods_supported: ["S256"],
token_endpoint_auth_methods_supported: ["none"],
});
}
token_endpoint_auth_method: "none" is correct and not a shortcut. An MCP
connector is a public client. It ships to end users and cannot keep a secret,
so PKCE, not a client secret, is what proves the token exchange came from the
same client that started the flow. (If you also serve confidential clients, say
a Zapier integration, add client_secret_basic and client_secret_post
alongside it.)
Dynamic client registration, honestly
RFC 7591 says a client can register itself. In practice, for a public client
authenticated by PKCE, there is nothing meaningful to store:
export async function POST(request: Request) {
const body = await request.json().catch(() => ({}));
return NextResponse.json({
client_id: "dfc_" + randomBytes(9).toString("base64url"),
client_name: body.client_name ?? "MCP Client",
redirect_uris: body.redirect_uris ?? [],
grant_types: ["authorization_code"],
response_types: ["code"],
token_endpoint_auth_method: "none",
});
}
That is the entire endpoint. It mints an id and hands it back. A clients table
would be ceremony: nothing downstream consults it.
But be careful about what you conclude from that. "We do not register clients,
PKCE covers it" is the sentence I would have written before I thought it
through, and it is wrong. See the next section.
The attack PKCE does not stop
PKCE binds an authorization code to whoever started the flow. The usual mental
model is that this makes an unregistered redirect_uri safe, because a
stolen code is useless without the verifier.
That model breaks when the attacker is the one who started the flow. They
craft an /authorize link with their own redirect_uri and their own
code_challenge, and send it to a signed-in victim. The victim's browser
follows it. The code is minted against the victim's session, redirects to the
attacker's callback, and the attacker exchanges it with the verifier they chose.
An access token for someone else's account, from one click. PKCE did its job
perfectly and protected nobody, because the attacker held the verifier all
along.
So the redirect cannot be automatic. GET /authorize renders a consent page
instead, and the code is only minted by a POST from that page:
export async function POST(request: Request) {
const origin = request.headers.get("origin");
if (origin !== url.origin) {
return NextResponse.json(
{ error: "invalid_request", error_description: "Cross-origin approval refused." },
{ status: 403 },
);
}
// ... re-read params, re-read the session, then mint
}
Two properties do the work. The approval is a step a crafted link cannot
perform on the victim's behalf. And Origin is sent on every form POST and
cannot be forged by page script, so a cross-site auto-submitting form is not an
approval either.
The consent page names the host the code is about to go to, not the full URI.
The host is the part that actually matters to the decision, and a long URI just
gives someone something to skim past.
Re-read the session inside the POST rather than trusting a hidden field. The
session is the only thing that says whose account the code is for.
One caveat on the redirect itself: use a 303, not a 307. Approval arrives as
a POST, and 307 preserves the method, so the browser would POST the code to a
callback that only answers GET. The client reports a bare "Bad Request" and it
looks like the client's bug rather than yours.
A confidential client with a pre-registered redirect_uri (a Zapier
integration, say) can skip the prompt. There is no third party to consent to,
because the caller could not have changed the destination.
PKCE
Twelve lines, all of it standard library:
import { randomBytes, createHash } from "crypto";
const CODE_TTL_MS = 5 * 60 * 1000;
export const generateAuthCode = () =>
"dfc_" + randomBytes(24).toString("base64url");
export const codeExpiry = () =>
new Date(Date.now() + CODE_TTL_MS).toISOString();
export function verifyPkce(verifier: string, challenge: string) {
return createHash("sha256").update(verifier).digest("base64url") === challenge;
}
Five-minute code TTL, single use, deleted on exchange. /authorize runs behind
your normal session check, so an unauthenticated user hits your existing login
first and comes back. That is the whole reason the user never sees a token.
If you also compare client secrets anywhere, use timingSafeEqual, and check
lengths yourself first. timingSafeEqual throws on a length mismatch, and an
uncaught throw becomes a 500 that leaks the secret's length.
4. Tokens and the caller
The token you issue can just be a row. Store a hash, never the token:
const { data: token } = await admin
.from("api_tokens")
.select("id, user_id, revoked_at, profiles(id, email, role, plan)")
.eq("token_hash", hashToken(raw))
.is("revoked_at", null)
.maybeSingle();
One query gets you validity and the caller's identity, role, and plan. The
connection then acts as that user, with their permissions, which means your
existing row-level security applies to MCP traffic for free. No parallel
permission model to keep in sync.
The part I got wrong first
I originally planned one endpoint with a capability flag, so a listed directory
connector could be restricted while my own stayed full-featured. It does not
work: an app directory registers its OAuth client against a base URL and then
freezes it. Whatever a listed URL is allowed to do has to be settled before
you list it.
So there are two endpoints, /api/mcp and /api/mcp/v2, same server behind
different options. The public one cannot see or create payment fields at all,
which means nothing built through the directory listing can collect money. Two
routes, five lines each:
export const { GET, POST, DELETE, OPTIONS } =
mcpHandlers({ path: "/api/mcp/v2", payments: false });
Capability decisions are URL-shaped, not runtime-shaped. Plan for that before
you submit anywhere.
Checklist
Stateless transport, sessionIdGenerator: undefined.
Server built per request with the caller baked in.
CORS on every response, including Expose-Headers.
401 carries WWW-Authenticate with the resource metadata URL.
Protected-resource metadata nested under each endpoint's own path.
Public client, PKCE, no client secret.
Registration endpoint can be trivial. Say why in a comment.
A consent page on /authorize, approved by same-origin POST. PKCE alone
does not stop a crafted-link attack, because there the attacker holds the
verifier.
303 on the post-approval redirect, not 307.
Token hashes in the database, RLS does the rest.
Decide capabilities per URL before you list anywhere.
None of this is much code. It is roughly 150 lines across six files. The hard
part was working out which pieces of the OAuth spec actually apply to a public
client that a user connects by pasting a URL, and which are ceremony.
The result is what I wanted: you paste
https://deoochform.com/api/mcp/v2 into Claude, a browser tab opens, you sign
in, and you are done. No token to copy, nothing to store. The
MCP server docs shows it from the user
side if you want to see what the flow feels like before building your own.