diff --git a/.changeset/sampling-context-capability-gate.md b/.changeset/sampling-context-capability-gate.md
new file mode 100644
index 0000000000..4390e78c84
--- /dev/null
+++ b/.changeset/sampling-context-capability-gate.md
@@ -0,0 +1,5 @@
+---
+'@modelcontextprotocol/server': patch
+---
+
+`Server.createMessage()` (and `ctx.mcpReq.requestSampling()`, which calls it) now refuses `includeContext: 'thisServer'` or `'allServers'` when the client did not declare the `sampling.context` capability, throwing `SdkError` with `SdkErrorCode.CapabilityNotSupported` before anything is sent, as it already does for `tools` without `sampling.tools`. The 2025-11-25 spec says servers SHOULD only use these values when the client declares `sampling.context`. A server that sends them to such a client must now omit `includeContext` or use `'none'`. Connections negotiated at a revision before 2025-11-25, which has no `sampling.context` capability, are unaffected.
diff --git a/packages/server/src/server/server.ts b/packages/server/src/server/server.ts
index 441857f646..e882d6f3c4 100644
--- a/packages/server/src/server/server.ts
+++ b/packages/server/src/server/server.ts
@@ -1107,6 +1107,15 @@ export class Server extends Protocol<ServerContext> {
         if ((params.tools || params.toolChoice) && !this._clientCapabilities?.sampling?.tools) {
             throw new SdkError(SdkErrorCode.CapabilityNotSupported, 'Client does not support sampling tools capability.');
         }
+        // sampling.context is a 2025-11-25 capability: older revisions allow thisServer/allServers without it
+        const predatesContextCapability = this._negotiatedProtocolVersion !== undefined && this._negotiatedProtocolVersion < '2025-11-25';
+        if (
+            (params.includeContext === 'thisServer' || params.includeContext === 'allServers') &&
+            !predatesContextCapability &&
+            !this._clientCapabilities?.sampling?.context
+        ) {
+            throw new SdkError(SdkErrorCode.CapabilityNotSupported, 'Client does not support sampling context capability.');
+        }
 
         // Message structure validation - always validate tool_use/tool_result pairs.
         // These may appear even without tools/toolChoice in the current request when
diff --git a/test/e2e/requirements.ts b/test/e2e/requirements.ts
index 36da4fc523..0e466235a1 100644
--- a/test/e2e/requirements.ts
+++ b/test/e2e/requirements.ts
@@ -1113,12 +1113,7 @@ export const REQUIREMENTS: Record<string, Requirement> = {
         behavior:
             'The server does not use includeContext values thisServer or allServers unless the client declared the sampling.context capability.',
         transports: ['inMemory', 'stdio', 'streamableHttp'],
-        note: 'Stateless hosting creates a fresh server per request and has no standalone GET stream, so there is no server→client channel to deliver/observe these.',
-        knownFailures: [
-            {
-                note: "Server.createMessage (src/server/index.ts ~line 497) only gates tools/toolChoice on clientCapabilities.sampling.tools; there is no check of clientCapabilities.sampling.context for includeContext 'thisServer'/'allServers', so the request reaches the client callback instead of being refused."
-            }
-        ]
+        note: 'Stateless hosting creates a fresh server per request and has no standalone GET stream, so there is no server→client channel to deliver/observe these.'
     },
     'sampling:create:not-supported': {
         source: 'https://modelcontextprotocol.io/specification/2025-11-25/client/sampling#capabilities',
diff --git a/test/e2e/scenarios/sampling.test.ts b/test/e2e/scenarios/sampling.test.ts
index c8048aad3b..004b8b602b 100644
--- a/test/e2e/scenarios/sampling.test.ts
+++ b/test/e2e/scenarios/sampling.test.ts
@@ -107,7 +107,8 @@ verifies('sampling:create:basic', async ({ transport }: TestArgs) => {
 
 verifies('sampling:create:include-context', async ({ transport }: TestArgs) => {
     const received: CreateMessageRequest[] = [];
-    const client = newClient();
+    // thisServer and allServers are only sent to a client that declares sampling.context.
+    const client = newClient({ sampling: { context: {} } });
     client.setRequestHandler('sampling/createMessage', async req => {
         received.push(req);
         return { model: 'm', role: 'assistant', stopReason: 'endTurn', content: { type: 'text', text: 'ok' } };
diff --git a/test/integration/test/server.test.ts b/test/integration/test/server.test.ts
index 0d5faabbe1..4fff0c72a3 100644
--- a/test/integration/test/server.test.ts
+++ b/test/integration/test/server.test.ts
@@ -1540,6 +1540,60 @@ describe('createMessage validation', () => {
         ).rejects.toThrow('Client does not support sampling tools capability.');
     });
 
+    test('should throw when includeContext thisServer or allServers is used without sampling.context capability', async () => {
+        const server = new Server({ name: 'test server', version: '1.0' }, { capabilities: {} });
+
+        const client = new Client(
+            { name: 'test client', version: '1.0' },
+            { capabilities: { sampling: {} } } // No context capability
+        );
+
+        client.setRequestHandler('sampling/createMessage', async () => ({
+            model: 'test-model',
+            role: 'assistant',
+            content: { type: 'text', text: 'Response' }
+        }));
+
+        const [clientTransport, serverTransport] = InMemoryTransport.createLinkedPair();
+        await Promise.all([client.connect(clientTransport), server.connect(serverTransport)]);
+
+        for (const includeContext of ['thisServer', 'allServers'] as const) {
+            await expect(
+                server.createMessage({
+                    messages: [{ role: 'user', content: { type: 'text', text: 'hello' } }],
+                    maxTokens: 100,
+                    includeContext
+                })
+            ).rejects.toThrow('Client does not support sampling context capability.');
+        }
+    });
+
+    test('should allow includeContext thisServer on a 2025-06-18 connection, which predates sampling.context', async () => {
+        const server = new Server({ name: 'test server', version: '1.0' }, { capabilities: {} });
+
+        const client = new Client(
+            { name: 'test client', version: '1.0' },
+            { capabilities: { sampling: {} }, supportedProtocolVersions: ['2025-06-18'] }
+        );
+
+        client.setRequestHandler('sampling/createMessage', async request => ({
+            model: 'test-model',
+            role: 'assistant',
+            content: { type: 'text', text: `includeContext=${request.params.includeContext}` }
+        }));
+
+        const [clientTransport, serverTransport] = InMemoryTransport.createLinkedPair();
+        await Promise.all([client.connect(clientTransport), server.connect(serverTransport)]);
+        expect(client.getNegotiatedProtocolVersion()).toBe('2025-06-18');
+
+        const result = await server.createMessage({
+            messages: [{ role: 'user', content: { type: 'text', text: 'hello' } }],
+            maxTokens: 100,
+            includeContext: 'thisServer'
+        });
+        expect(result.content).toEqual({ type: 'text', text: 'includeContext=thisServer' });
+    });
+
     test('should throw when tool_result is mixed with other content', async () => {
         const server = new Server({ name: 'test server', version: '1.0' }, { capabilities: {} });