Autograd — MCP Tool Specification
Version: v1.1.9 | Status: Active | Last Updated: March 2026
Overview
This document defines the MCP tools exposed by the autograd module.
These tools are auto-discovered by the PAI MCP bridge via the @mcp_tool decorator
in mcp_tools.py and surfaced as part of the ~303 dynamic tools available to Claude.
The autograd module provides automatic differentiation capabilities, allowing AI agents to evaluate mathematical expressions and verify gradient computations.
Auto-Discovery
| Property | Value |
|---|---|
| Discovery method | @mcp_tool decorator scan |
| Namespace | autograd |
| Trust default | Safe |
| PAI bridge | src/codomyrmex/agents/pai/mcp/ |
Tool Reference
autograd_compute
Description: Evaluate a simple expression and compute its gradient via reverse-mode automatic differentiation. Trust Level: Safe Category: analysis
Parameters:
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
expression | str | Yes | -- | Simple math expression like "x*x + y" using +, -, *, ** operators |
variables | dict | Yes | -- | Dict of variable names to float values, e.g. {"x": 2.0, "y": 3.0} |
Returns: dict — Dictionary with result (float, the expression value) and gradients (dict mapping variable names to their gradient values).
Example:
from codomyrmex.autograd.mcp_tools import autograd_compute
result = autograd_compute(expression="x*x + y", variables={"x": 3.0, "y": 1.0})
# {"result": 10.0, "gradients": {"x": 6.0, "y": 1.0}}
Notes: Uses a restricted eval with no builtins. Only supports +, -, *, ** operators and variable names that are valid Python identifiers.
autograd_gradient_check
Description: Numerically verify that analytic gradients match finite differences. Trust Level: Safe Category: analysis
Parameters:
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
func_name | str | Yes | -- | One of "relu", "tanh", "sigmoid", "square", "sum" |
inputs | list | Yes | -- | List of input values to check gradients at |
Returns: dict — Dictionary with max_error (float), passed (bool, true if max_error < 1e-4), analytic_grads (list), and numeric_grads (list).
Example:
from codomyrmex.autograd.mcp_tools import autograd_gradient_check
result = autograd_gradient_check(func_name="sigmoid", inputs=[0.0, 1.0, -1.0])
# {"max_error": 1.2e-10, "passed": true, ...}
Notes: Uses central finite differences with epsilon=1e-5. Supported functions: relu, tanh, sigmoid, square, sum.
Integration Notes
- Auto-discovered: Yes (via
@mcp_toolinmcp_tools.py) - Trust Gateway: All tools are safe — no trust check required
- PAI Phases: VERIFY (gradient correctness validation), BUILD (expression evaluation)
- Dependencies:
autograd.engine.Value,autograd.ops(relu, sigmoid, tanh)