Model and verify bounded LP and MILP problems with NVIDIA cuOpt
Use NVIDIA’s portable optimization skill to formulate a bounded LP or MILP, run cuOpt, and check solver status and assignments against the intended constraints.
npx skills add agentskillexchange/skills --skill model-and-verify-bounded-lp-and-milp-problems-with-nvidia-cuopt
Use NVIDIA’s cuopt-numerical-optimization-api skill when a coding agent must turn a bounded allocation, capacity or production-planning problem into a reviewable LP or MILP model and solver result. Supply the objective, coefficients, constraints, units, variable bounds, integrality requirements and allowed solve time. The agent confirms the interface, models the problem, runs cuOpt and reports solver status, objective and variable assignments against the original constraints.
What this skill actually does
Invoke this when the work requires translating operator requirements into a solver model or checking an existing model from an agent session. For an existing MPS or LP file, use the documented CLI interface; for a new model from data, prefer the Python API or an existing supported modeling tool rather than hand-writing a model file.
The boundary is formulation, bounded execution and validation of one optimization problem. This entry covers the named upstream skill rather than the NVIDIA skills catalog as a whole, general CUDA development, vehicle-routing APIs or automatic operational decisions. Keep quadratic optimization outside the default workflow because upstream labels it beta.
Workflow: confirm problem type and interface; check GPU/CUDA prerequisites; define continuous versus integer or binary variables, objective, constraints and bounds; set a solve deadline and appropriate MILP gap; solve; inspect the actual documented status; distinguish an optimal solution from a merely feasible incumbent or time-limited result; check feasibility and units before handing the result to the operator. Report infeasible, unbounded or numerical-error results without inventing assignments.
Python invocation uses cuopt.linear_programming.problem.Problem and cuopt.linear_programming.solver_settings.SolverSettings. The upstream Python reference demonstrates addVariable, addConstraint, setObjective, settings.set_parameter and problem.solve(settings). Status names use PascalCase, including Optimal and FeasibleFound; use the reference appropriate to the installed version. The CLI entry point is cuopt_cli <problem-file>; inspect cuopt_cli –help for supported settings.
Installation: Install the Agent Skills package with npx skills@latest add nvidia/skills –skill cuopt-numerical-optimization-api –agent claude-code –yes; use –agent codex or another supported target when appropriate. Confirm the installed skill and load it in the target agent. Provision cuOpt separately using the upstream cuopt-install instructions for the actual GPU, driver and CUDA version; do not mix CUDA package variants. Verify the Python import or cuopt_cli –help before solving.
Framework fit: Multi-Framework. NVIDIA documents the same portable skill for Claude Code, Codex, Cursor and other Agent Skills hosts; the solver interface is independent of the host. Attribution: NVIDIA cuOpt Team. The inspected named skill declares Apache-2.0. Repository stars describe adoption of NVIDIA/skills, not separate usage of this individual skill. Metadata was checked; installation, GPU execution and security review were not performed.