Ortools

An MCP server that exposes Google OR-Tools CP-SAT as tools for constraint-based value correction. Given a set of equations and observed variable values, it finds the minimal corrections needed to make everything consistent.

Use case: you have values extracted from a document (OCR, form parsing, etc.) that should satisfy known equations (e.g. total = quantity × rate + tax), but one or more values were misread. The solver finds which values to adjust and by how much, weighted by how much you trust each observation.

---

Tools

solve

Runs CP-SAT. Returns corrected values for free (non-fixed) variables that best satisfy the given equations.

Input

{
  "variables": {
    "gross_amount": {
      "obs": 500,
      "confidence": 0.3,
      "mult_factor": 100,
      "min": 0,
      "max": 100000,
      "fixed": false
    },
    "quantity": {
      "obs": 10,
      "mult_factor": 1,
      "fixed": true
    },
    "rate_per_pc": {
      "obs": 52,
      "mult_factor": 100,
      "fixed": true
    }
  },
  "equations": [
    {
      "lhs": "gross_amount",
      "rhs": "quantity * rate_per_pc",
      "relation": "==",
      "weight": 1000,
      "tolerance": 0
    }
  ],
  "timeout_seconds": 5,
  "num_workers": 8
}

Output

{
  "status": "OPTIMAL",
  "corrected": {
    "gross_amount": 520.0
  }
}

Only free (non-fixed) variables appear in corrected. Status values: OPTIMAL, FEASIBLE, INFEASIBLE, UNKNOWN, INVALID_EQUATION.

Variable spec fields

Field	Type	Default	Description
obs	number	—	Observed (extracted) value in real-world units
confidence	0–1	0.0	How much to trust obs. Higher → stronger pull toward the observed value
mult_factor	int	1	Scales float values to integers for the solver (obs × mult_factor), result is divided back on return
min / max	number	0 / 2e9	Domain bounds in real-world units
fixed	bool	false	Pin this variable to obs — it won't appear in corrected

Equation fields

Field	Type	Default	Description
lhs	string	required	Left-hand side expression
rhs	string	required	Right-hand side expression
relation	== | <= | >=	required	Relation between lhs and rhs
weight	int	1000	Penalty per unit of violation — higher means the solver tries harder to satisfy this equation
tolerance	int	0	Allowed slack (in scaled integer units) before penalty kicks in

Solver params

Field	Default	Description
timeout_seconds	5	Wall-clock limit for the solver
num_workers	8	Parallel search workers

---

validate_equations

Pure AST check — no solver invoked. Call this before solve to catch unsupported constructs early.

Input

{
  "equations": [
    { "lhs": "total", "rhs": "taxable + tax" },
    { "lhs": "discount", "rhs": "-base * 0.1" }
  ]
}

Output

[
  { "lhs": "total", "rhs": "taxable + tax", "errors": [] },
  {
    "lhs": "discount",
    "rhs": "-base * 0.1",
    "errors": [
      "Unary operator 'USub' not supported in '-base * 0.1'. Rewrite e.g. '-x' as '(0 - x)'."
    ]
  }
]

Empty errors list means the equation is valid. Catches:

• Unary negation (-x) and unary plus (+x) — rewrite as (0 - x)
• Non-constant or zero/negative exponents in ** — only x ** 2, x ** 3, etc. are supported
• Syntax errors

---

evaluate_equations

Evaluates equations against known values and returns pass/fail per equation. No OR-Tools involved — useful to check whether values already satisfy the system before deciding to call solve.

Input

{
  "equations": [
    { "lhs": "gross_amount", "rhs": "quantity * rate_per_pc", "relation": "==", "tolerance": 0.5 }
  ],
  "values": {
    "gross_amount": 520,
    "quantity": 10,
    "rate_per_pc": 52
  }
}

Output

[
  {
    "lhs": "gross_amount",
    "rhs": "quantity * rate_per_pc",
    "status": "passed",
    "actual": 520,
    "computed": 520,
    "error": 0.0
  }
]

Status values: passed, failed, missing_values (variable not in values), error (evaluation exception).

---

optimize

Maximize or minimize an objective expression subject to constraints. Unlike solve, there are no observed values — this is pure optimization: find the best possible assignment of variables within their domains.

Example use cases:

• Maximize margin (revenue - cost) subject to budget and inventory limits
• Minimize waste subject to production constraints
• Find the best allocation of a fixed total across items

Input

{
  "variables": {
    "units": { "min": 0, "max": 1000, "mult_factor": 1 },
    "price": { "min": 0, "max": 500,  "mult_factor": 100 }
  },
  "objective": {
    "expr": "units * price",
    "direction": "maximize"
  },
  "hard_constraints": [
    { "lhs": "units + price", "rhs": "120", "relation": "<=" }
  ],
  "soft_constraints": [
    { "lhs": "price", "rhs": "50", "relation": "<=", "weight": 500, "tolerance": 0 }
  ],
  "timeout_seconds": 5,
  "num_workers": 8
}

Output

{
  "status": "OPTIMAL",
  "values": { "units": 100, "price": 20.0 },
  "objective_value": 2000.0
}

• values contains all variables (real-world units, divided by mult_factor)
• objective_value is computed from the real-world values after solving
• Status values: OPTIMAL, FEASIBLE, INFEASIBLE, UNKNOWN, INVALID_EQUATION

Constraint types

Type	Behaviour
hard_constraints	Must be satisfied. If any are contradictory the solver returns INFEASIBLE.
soft_constraints	Violations are penalised in the objective (for maximize: subtracted; for minimize: added). Same weight/tolerance fields as solve.

Variable spec (simpler than solve — no obs, confidence, or fixed):

Field	Default	Note
min	0	Domain lower bound (real-world units)
max	2e9 // mult_factor	Domain upper bound (real-world units)
mult_factor	1	Same integer-scaling as solve

---

Supported expression syntax

Both lhs and rhs accept arithmetic expressions over variable names and numeric constants:

Operator	Example	Notes
+	a + b
-	a - b	Binary only — -a (unary) is not supported
*	qty * rate
/	total / qty	RHS divisions are automatically moved to LHS to avoid integer-division loss
%	a % b	Modulo
**	x ** 2	Exponent must be a constant positive integer
Parentheses	(a + b) * c

---

mult_factor and integer scaling

OR-Tools CP-SAT works only with integers. mult_factor scales real-world float values to integers before solving and divides back on return.

• Amount fields (prices, totals): mult_factor: 100 — values like 52.30 become 5230 internally
• Quantity fields: mult_factor: 1 — must stay at 1 for multiplicative equations

Why quantities must use mult_factor: 1: for gross = qty × rate, scaling all three by 100 gives gross×100 = (qty×100) × (rate×100), which is off by 100×. With qty at ×1: gross×100 = (qty×1) × (rate×100) — correct.

tolerance is expressed in scaled integer units (after mult_factor is applied). For mult_factor: 100, a tolerance of 200 means ±2.00 in real units.

---

Setup

pip install -r requirements.txt

Add to Claude Code (~/.claude/settings.json):

{
  "mcpServers": {
    "ortools-solver": {
      "command": "python3",
      "args": ["/path/to/ortools-mcp/server.py"]
    }
  }
}

---

Running tests

pip install pytest
pytest test_solver.py test_server.py -v

test_server.py mocks the mcp package so it runs without needing mcp installed. 50 tests total covering all three tools, expression validation, scaling, confidence weighting, tolerance, and inequality constraints.