numkit-mcp

MCP server exposing the numkit numerical computing engine as tools for AI agents.

Why

AI agents (Claude Desktop, Cursor, Claude Code, Continue, AutoGen, LangChain) need reliable numerical computation. Python/Jupyter sandboxes are heavy (300–800 ms cold start, ~500 MB) and token-expensive. numkit-mcp provides:

• Instant startup (< 1 ms after first load — in-process WASM)
• Persistent workspace — define a function in one call, use it in the next
• Token-efficient — native MATLAB matrix syntax (A \ b, fft(x), [B,A] = butter(4, 0.2))
• Zero dependencies — the WASM engine ships inside the package

Install

{
  "mcpServers": {
    "numkit": {
      "command": "npx",
      "args": ["-y", "numkit-mcp"]
    }
  }
}

Tools

| Tool | Purpose | |---|---| | numkit_eval(code) | Execute MATLAB/numkit code in a persistent workspace | | numkit_reset() | Clear the workspace | | numkit_help(fn) | Get function documentation |

Example session

Agent: numkit_eval("A = [1 2; 3 4]; b = [5; 11]; x = A \ b")
→ x = [-3; 4]

Agent: numkit_eval("eig(A)")
→ eigenvalues of A (workspace survived from the previous call)

Agent: numkit_eval("[B, A] = butter(4, 0.2); freqz(B, A)")
→ filter design + frequency response

Configuration

| Env var | Default | Description | |---|---|---| | NUMKIT_MCP_TIMEOUT | 3000 | Max execution time per call (ms) |

NOT a filesystem sandbox

Like any scripting-language CLI, scripts can read and write files in the working directory. Do not use this to execute untrusted code.

License

0BSD (same as numkit).