A method for spatial deconvolution of spectra is presented. It follows the same fundamental principles as the ``MCS image deconvolution algorithm'' (Magain, Courbin, Sohy, 1998) and uses information contained in the spectrum of a reference Point Spread Function (PSF) to spatially deconvolve spectra of very blended sources. An improved resolution rather than an infinite one is aimed at, overcoming the well known problem of ``deconvolution artefacts''. As in the MCS algorithm, the data are decomposed into a sum of analytical point sources and a numerically deconvolved background, so that the spectrum of extended sources in the immediate vicinity of bright point sources may be accurately extracted and sharpened. The algorithm has been tested on simulated data including seeing variation as a function of wavelength and atmospheric refraction. It is shown that the spectra of severely blended point sources can be resolved while fully preserving the spectrophotometric properties of the data. Extended objects ``hidden'' by bright point sources (up to 4-5 magnitudes brighter) can be accurately recovered as well, provided the data have a sufficiently high total signal-to-noise ratio (200-300 per spectral resolution element). Such spectra are relatively easy to obtain, even down to faint magnitudes, within a few hours of integration time with 10m class telescopes.