NumPy Computation Server

Executes NumPy operations through the Model Context Protocol (MCP) transport layer, translating LLM-generated function calls into native Python NumPy operations with automatic type marshalling and error handling. Uses MCP's resource and tool abstractions to expose NumPy's 200+ array manipulation functions as callable tools with JSON schema validation, enabling stateless computation requests from any MCP-compatible client.

Provides high-level array operations (reshape, transpose, concatenate, split, slice) with automatic dimension inference and broadcasting rules applied transparently. Translates user intent (e.g., 'flatten this 3D array') into optimized NumPy calls, handling edge cases like mismatched dimensions or incompatible shapes with descriptive error messages that guide users toward valid operations.

Exposes NumPy's logical and comparison functions (equal, not_equal, greater, less, logical_and, logical_or, logical_not) through MCP tool definitions with automatic broadcasting and dtype handling. Implements element-wise comparisons that return boolean arrays suitable for masking or conditional operations.

Exposes NumPy's dtype conversion and inspection functions (astype, dtype property, itemsize, nbytes) through MCP tool definitions with automatic dtype validation and conversion safety checks. Implements schema-based dtype specification supporting all NumPy scalar types (int, float, complex, bool, string, datetime).

Exposes NumPy's linear algebra module (numpy.linalg) through MCP, enabling matrix operations like inversion, eigenvalue decomposition, SVD, QR factorization, and solving linear systems. Handles numerical stability concerns (singular matrices, ill-conditioned systems) with fallback strategies and returns decomposition results with condition number warnings when appropriate.

Implements statistical functions (mean, median, std, percentile, correlation, covariance) with axis-aware reduction, allowing aggregation along specific dimensions while preserving others. Automatically handles NaN/Inf values with configurable strategies (skip, fill, propagate) and returns both point estimates and confidence intervals when requested.

Exposes NumPy's random module (numpy.random) through MCP with support for multiple distributions (normal, uniform, exponential, Poisson, etc.) and explicit seed control for reproducible results. Maintains separate random state per request or allows persistent state management for sequential sampling workflows.

Provides polynomial fitting (numpy.polyfit) and evaluation (numpy.polyval) capabilities, allowing LLM agents to fit polynomial curves to data and evaluate them at new points. Includes automatic degree selection heuristics and residual analysis to help users choose appropriate polynomial orders without overfitting.

Exposes NumPy's FFT module (numpy.fft) through MCP, enabling 1D and 2D Fourier transforms, inverse transforms, and frequency domain analysis. Automatically handles real vs. complex input, applies windowing functions to reduce spectral leakage, and returns frequency bins with magnitude/phase information.

Provides sorting (argsort, sort) and searching (searchsorted, where) capabilities with support for multi-key sorting and custom axis specification. Enables LLM agents to sort arrays by multiple columns, find insertion points for binary search, and filter arrays based on conditions without explicit loop construction.

Implements element-wise mathematical functions (sin, cos, exp, log, sqrt, etc.) with automatic NumPy broadcasting, allowing operations on arrays of different shapes that are compatible under broadcasting rules. Handles edge cases (division by zero, log of negative numbers) with configurable error handling (raise, warn, or return NaN).

Provides logical operations (AND, OR, NOT) and comparison operations (==, <, >, etc.) with automatic broadcasting and boolean array generation. Enables complex filtering through boolean indexing, allowing LLM agents to extract elements matching multiple conditions without explicit loop construction.