P2P GPU matrix multiply, exact cosine search, and PCA/SVD verified via Freivalds' algorithm.
Maintain this server? Claim it
Call this server's tools from your browser. Nothing to install.
Requests go through mcptry because browsers can't call most MCP servers directly. Header values are sent to the server for that request only and are never stored or logged.
Tool list from our last check. Connect live to refresh it.
Runs exact cosine similarity search between query embeddings (M x D) and corpus embeddings (N x D) on distributed GPUs using OUTPUT_GRID_2D zero-reduction tiling.
File path (.npy/.npz/.safetensors) or inline 2D array of shape (M, D).
File path (.npy/.npz/.safetensors) or inline 2D array of shape (N, D).
Target GEMM precision.
Number of top matching corpus indices and scores to return per query.
Optional file path (.npy/.safetensors) to save the full similarity matrix.
This tool isn't marked read-only, so it may change something on the server or in an account you connected.
Pick your client. Values in capitals, like YOUR_TOKEN, are yours to fill in.
Run in a terminal
claude mcp add --transport http matrix-compute-mcp https://mcp.matrix-compute.net/mcp
Runs exact cosine similarity search between query embeddings (M x D) and corpus embeddings (N x D) on distributed GPUs using OUTPUT_GRID_2D zero-reduction tiling.
| Argument | Type | Description |
|---|---|---|
| dtype | string | Target GEMM precision. |
| top_k | integer | Number of top matching corpus indices and scores to return per query. |
| queries* | any | File path (.npy/.npz/.safetensors) or inline 2D array of shape (M, D). |
| output_path | string | Optional file path (.npy/.safetensors) to save the full similarity matrix. |
| corpus_embeddings* | any | File path (.npy/.npz/.safetensors) or inline 2D array of shape (N, D). |
{
"type": "object",
"required": [
"queries",
"corpus_embeddings"
],
"properties": {
"dtype": {
"enum": [
"fp32",
"fp16",
"bf16"
],
"type": "string",
"default": "fp32",
"description": "Target GEMM precision."
},
"top_k": {
"type": "integer",
"default": 5,
"description": "Number of top matching corpus indices and scores to return per query."
},
"queries": {
"description": "File path (.npy/.npz/.safetensors) or inline 2D array of shape (M, D)."
},
"output_path": {
"type": "string",
"description": "Optional file path (.npy/.safetensors) to save the full similarity matrix."
},
"corpus_embeddings": {
"description": "File path (.npy/.npz/.safetensors) or inline 2D array of shape (N, D)."
}
},
"additionalProperties": false
}Computes verified top-k Principal Component Analysis (PCA) / Singular Value Decomposition (SVD) on distributed GPU workers via randomized subspace iteration or stateful power iteration.
| Argument | Type | Description |
|---|---|---|
| k* | integer | Number of top singular/principal components to extract. |
| center | boolean | Mean-center columns before decomposition (standard PCA). |
| matrix* | any | File path (.npy/.npz/.safetensors) or inline 2D array of shape (M, N). |
| method | string | Decomposition algorithm ('randomized' or 'power_iteration'). |
| output_path | string | Optional output file path (.npz or .safetensors) to save U, S, Vt. |
{
"type": "object",
"required": [
"matrix",
"k"
],
"properties": {
"k": {
"type": "integer",
"description": "Number of top singular/principal components to extract."
},
"center": {
"type": "boolean",
"default": true,
"description": "Mean-center columns before decomposition (standard PCA)."
},
"matrix": {
"description": "File path (.npy/.npz/.safetensors) or inline 2D array of shape (M, N)."
},
"method": {
"enum": [
"randomized",
"power_iteration"
],
"type": "string",
"default": "randomized",
"description": "Decomposition algorithm ('randomized' or 'power_iteration')."
},
"output_path": {
"type": "string",
"description": "Optional output file path (.npz or .safetensors) to save U, S, Vt."
}
},
"additionalProperties": false
}Calculates exact upload/download network bandwidth (MB), task count, GFLOPs, and satoshi cost estimates for a distributed matrix multiplication C (m x n) = A (m x k) x B (k x n).
| Argument | Type | Description |
|---|---|---|
| k* | integer | Inner dimension K (columns of A / rows of B). |
| m* | integer | Rows of matrix A (M > 0). |
| n* | integer | Columns of matrix B (N > 0). |
| dtype | string | Precision format ('fp32', 'fp16', 'bf16'). |
| grid_cols | integer | Column partitions P_c when using output_grid_2d. |
| grid_rows | integer | Row partitions P_r when using output_grid_2d. |
| tile_size | integer | Tile block size dimension. |
| num_strips | integer | Strip count when using outer_product_1d. |
| sats_per_gflop | number | Override worker price in satoshis per GFLOP. |
| partitioning_strategy |
Executes verified distributed matrix multiplication C = A x B across decentralized GPU workers with dynamic Freivalds verification and L402 settlement. Accepts file paths (.npy, .npz, .safetensors) or inline 2D JSON arrays.
| Argument | Type | Description |
|---|---|---|
| dtype | string | Computation precision ('fp32', 'fp16', 'bf16'). |
| matrix_a* | any | File path (.npy/.npz/.safetensors) or inline 2D array for matrix A (M x K). |
| matrix_b* | any | File path (.npy/.npz/.safetensors) or inline 2D array for matrix B (K x N). |
| tile_size | integer | Optional tile block size override. |
| output_path | string | Optional output file path (.npy, .npz, or .safetensors) to write C. |
| tensor_key_a | string | Optional tensor key when loading matrix_a from .safetensors or .npz. |
| tensor_key_b | string | Optional tensor key when loading matrix_b from .safetensors or .npz. |
| partitioning_strategy | string | Matrix partitioning topology. |
Queries the Matrix Compute discovery relay for live GPU worker availability, total VRAM capacity, confidential computing (TEE) enclaves, and sats_per_gflop pricing.
| Argument | Type | Description |
|---|---|---|
| min_vram_mb | integer | Minimum worker VRAM filter in MB (optional). |
| max_price_per_gflop | number | Maximum worker price filter in satoshis per GFLOP (optional). |
{
"type": "object",
"properties": {
"min_vram_mb": {
"type": "integer",
"description": "Minimum worker VRAM filter in MB (optional)."
},
"max_price_per_gflop": {
"type": "number",
"description": "Maximum worker price filter in satoshis per GFLOP (optional)."
}
},
"additionalProperties": false
}We read every tool's name, description and inputs the way the model will, looking for instructions hidden from you. A clean result means no warning signs, not a guarantee. Scanned Nov 15, 2026.
No readOnlyHint / destructiveHint annotations, so clients can't tell a read from a write without asking.
Fix: Add annotations to each tool, e.g. { readOnlyHint: true } for tools that only read.
| string |
| Matrix partitioning strategy. |
{
"type": "object",
"required": [
"m",
"k",
"n"
],
"properties": {
"k": {
"type": "integer",
"description": "Inner dimension K (columns of A / rows of B)."
},
"m": {
"type": "integer",
"description": "Rows of matrix A (M > 0)."
},
"n": {
"type": "integer",
"description": "Columns of matrix B (N > 0)."
},
"dtype": {
"enum": [
"fp32",
"fp16",
"bf16"
],
"type": "string",
"default": "fp32",
"description": "Precision format ('fp32', 'fp16', 'bf16')."
},
"grid_cols": {
"type": "integer",
"description": "Column partitions P_c when using output_grid_2d."
},
"grid_rows": {
"type": "integer",
"description": "Row partitions P_r when using output_grid_2d."
},
"tile_size": {
"type": "integer",
"default": 1024,
"description": "Tile block size dimension."
},
"num_strips": {
"type": "integer",
"default": 4,
"description": "Strip count when using outer_product_1d."
},
"sats_per_gflop": {
"type": "number",
"description": "Override worker price in satoshis per GFLOP."
},
"partitioning_strategy": {
"enum": [
"output_grid_2d",
"outer_product_1d",
"block_3d"
],
"type": "string",
"default": "output_grid_2d",
"description": "Matrix partitioning strategy."
}
},
"additionalProperties": false
}{
"type": "object",
"required": [
"matrix_a",
"matrix_b"
],
"properties": {
"dtype": {
"enum": [
"fp32",
"fp16",
"bf16"
],
"type": "string",
"default": "fp32",
"description": "Computation precision ('fp32', 'fp16', 'bf16')."
},
"matrix_a": {
"description": "File path (.npy/.npz/.safetensors) or inline 2D array for matrix A (M x K)."
},
"matrix_b": {
"description": "File path (.npy/.npz/.safetensors) or inline 2D array for matrix B (K x N)."
},
"tile_size": {
"type": "integer",
"description": "Optional tile block size override."
},
"output_path": {
"type": "string",
"description": "Optional output file path (.npy, .npz, or .safetensors) to write C."
},
"tensor_key_a": {
"type": "string",
"description": "Optional tensor key when loading matrix_a from .safetensors or .npz."
},
"tensor_key_b": {
"type": "string",
"description": "Optional tensor key when loading matrix_b from .safetensors or .npz."
},
"partitioning_strategy": {
"enum": [
"output_grid_2d",
"outer_product_1d",
"block_3d"
],
"type": "string",
"default": "output_grid_2d",
"description": "Matrix partitioning topology."
}
},
"additionalProperties": false
}