When an observation is given at a sequence of positions along the fluid flow, then it can be defined as Lagrangian, in a mathematical point of view. From this sequence of positions (for instance the profiling drifting floats of Argo program), one can deduce important information on the stream that transports the drifters. Such an information has not yet been exploited in an operational framework, although previous works \cite{nodet:06} have shown the interest of assimilating this new type of data. A task of the ANR VODA has thus been defined in order to develop the tools for the variationnal assimilation of Lagrangian data in the context of NEMOVAR. We present here the writing of the lagrangian observation operator, and the construction of the associated tangent and adjoint models. The observation operator requires the interpolation of the velocity at any point of the domain. This interpolation operator is not linear for the general grids used in NEMO, implying heavy tangent and adjoint operators. Tests of these procedures are done on the main test configurations GYRE and ORCA2 . One aim of the development of these routines was to make the assimilation of Lagrangian Data possible for users of NEMOVAR. This has then required an implementation in the code with the respect to the philosophy of the other observation operators, in term of fortran data structures as well as interface with the user.