As @tylo says, projecting the point to the $x$ coordinate does not give you a homomorphism. So
this version is not useful if you want to have additively homomorphic ElGamal.
However, you could use the "exponential" version of standard ElGamal on elliptic
curves, i.e., instead of encrypting a message $m$ somehow mapped to a point $M$ on the curve (using an injective efficiently invertible encoding), to encrypt a message $m$ straightforwardly mapped as $M=m\cdot P$ to a point on the curve where $P$ is your generator and $m$ an integer in the order
of the group. This will give you an additive homomorphic encryption
scheme. However, as discussed below, this encoding is not efficiently invertible.
Let $P$ be the generator of your elliptic curve group of prime order $q$ and $Y=xP$ be your
public key ($x$ the private key). Then given two ciphertexts for messages $m_1,m_2\in Z_q$:
$C_1=(k_1P,m_1P+k_1Y)$ and $C_2=(k_2P,m_2P+k_2Y)$, then by componentwise point addition you receive
$$C=((k_1+k_2)P,(m_1+m_2)P+(k_1+k_2)Y)$$ which is a valid ciphertext for message $m_1+m_2 \mod q$.
You can decrypt $C$, but this obviously gives you $M=(m_1+m_2)P$ and in order to recover
$m_1+m_2 \mod q$, you have to compute $\log_P M$, i.e., you have to compute discrete logarithms after
encryption. If your messages $m_i$ come from a small set, this, however, is quite efficiently feasible. However,
this clearly depends on your application, i.e., which values from which range the $m_i$'s can take.