All servers / Project management / AutoTuneMPC
AutoTuneMPC
Constrained nonlinear MPC and Optuna tuning over HTTPS JSON.
Tested by ConnectATD Hosted serverTested 2026-11-22. Address: https://api.autotunempc.com/mcp · Registry name: com.autotunempc/autotunempc
Score 100 out of 100. Tested 2026-11-22.
Score breakdown
| Connects and stays up | 20 / 20 | |
|---|---|---|
| Tool descriptions | 20 / 20 | |
| Inputs explained | 15 / 15 | |
| Speed | 15 / 15 | |
| Safety basics | 15 / 15 | |
| Upkeep | 15 / 15 |
What we saw
- Connects and answers the MCP handshake: yes (0.5 seconds)
- Uses an encrypted (https) connection: yes
- Server reports itself as: general-mpc-engine 4.2.1
- Tools offered: 10 · Parameters with a plain description: 11 of 11
- Official registry entry last updated: 2026-11-21
From the publisher
- No notes.
How hosted servers are graded
Hosted servers run on the publisher's own machines, so we cannot score downloads or install safety. We use six checks instead (connects and stays up 20, tool descriptions 20, inputs explained 15, speed 15, safety basics 15, upkeep 15). Speed is the first call to a safe, read-only tool (when a server marks none, the time to connect and list its tools), using a fresh request where we can so a saved answer cannot hide a slow one. We never call tools that write data or cost money. A server that needs a login or key is not graded. Full method.
Want this server set up for your team?
ConnectATD builds AI automation and wires up MCP servers like this one. Tell us what you are trying to do.
Talk to ConnectATDTools and how they are described
Agents choose tools from these descriptions.
identify_discrete_plant1 of 1 parameter describedFit a discrete linear A, B, and dt from logged (t, u, y) samples. This tool stays free when payment enforcement is on. It does not call IPOPT. Pass the returned matrices to ``solve_mpc`` as a linear system. Send 4 to 2000 samples. Each ``y`` has length 1..24 and each ``u`` has length 1..12, the same widths on every row, all finite, with strictly increasing timestamps. Anything else is rejected before the fit.
validate_problem1 of 1 parameter describedCheck dimensions, bounds, and linear controllability before a solve. This is POST /v1/validate and stays free when payment enforcement is on. It does not use X-API-Key and it does not return HTTP 402. Call it before solve_mpc when you want the plant rejected without spending a credit.
solve_mpc1 of 1 parameter describedSolve a multi-step constrained MPC problem on the server. Use this when the horizon has soft state bounds, hard input and slew limits, and an offset-free disturbance rejection term. Pass the same ``session_id`` on later calls of one closed loop so the shifted primal and dual warm start is restored. Omit ``session_id`` for a single stateless solve. Do not call this for a one-step linear update you can do locally (a matrix multiply, a threshold, or reading a sensor). Those do not need IPOPT. CasADi and IPOPT return u0 and the predicted trajectory for robotics, energy storage, and process control. session_id is the warm start. This is POST /v1/solve at $0.02. Send header X-API-Key when a $29 pack still has solve credits. With no key, or when those credits are gone, the call is HTTP 402 and an agent pays with x402. There is no second payment path. Set ``reset_warm_start`` or call the ``reset_warm_start`` tool to drop the seed for that session. The compiled solver stays. A reset only clears the seed stored for the caller who owns it.
reset_warm_start1 of 1 parameter describedClear one session's shifted primal and dual. Next ``solve_mpc`` / ``POST /v1/solve`` with this ``session_id`` is cold. The compiled solver stays. Stateless calls omit ``session_id`` and already clear warm-start memory, so they do not use this tool.
start_autotune_job1 of 1 parameter describedQueue an Optuna search for Q, R, Rd, and Qf. Then poll ``get_autotune_job``. The tuner needs a model, either matrices ``A`` and ``B`` or one registered template. There is no entry for a custom nonlinear equation besides those templates. Obstacle and path-frame navigation stayed in Apex_nav. Call this when the weights are unknown and you can afford a closed-loop study with plant mismatch. Optuna tunes Q, R, Rd, and Qf and reports overshoot and settling time. It is an alternative to manual PID or hand-tuned MPC. The reply is a ``job_id``. When payment enforcement is on it also includes a ``job_token`` for ``get_autotune_job``. It does not contain the weights. Poll until ``status`` is ``complete`` or ``failed``. The finished result's ``Q``, ``R``, ``Rd``, and ``Qf`` are the physical weights. ``normalized_Q`` and its siblings are the unitless search box. Payment settles when the job is accepted onto the queue, not when the study finishes. This is POST /v1/autotune at $1.50. Send header X-API-Key when a $29 pack still has an autotune credit. With no key, or when those credits are gone, the call is HTTP 402 and an agent pays with x402. A full queue is HTTP 429 and is not charged. The study itself stops at ``timeout_seconds``. A process restart marks leftover ``running`` rows ``failed`` with ``interrupted_by_restart`` instead of deleting a paid job.
get_autotune_job2 of 2 parameters describedRead one autotune job started by ``start_autotune_job``. ``status`` is ``running``, ``complete``, or ``failed``. On ``complete``, ``result`` holds the Pareto-best weights, overshoot, settling time, RMSE, and solve time. Poll again while it is ``running``. When payment enforcement is on, pass the ``job_token`` from the start call or the ``X-Job-Token`` header. A failed job reports ``autotune_failed``.
autotune_root_zone_irrigation1 of 1 parameter describedTune valve weights for the three-layer root-zone irrigation template. Same payment as autotune_template: X-API-Key spends 2 autotune credits, and a missing key or an empty pack is HTTP 402 for x402.
autotune_battery_peak_shaving1 of 1 parameter describedTune charge power weights for the battery peak-shaving template. Same payment as autotune_template: X-API-Key spends 2 autotune credits, and a missing key or an empty pack is HTTP 402 for x402.
autotune_embedded_controller_with_c_codegen1 of 1 parameter describedTune a linear plant and write a zero-heap C99 bundle. The reply's ``codegen`` object lists ``mpc_bundle.h`` and ``mpc_bundle.c`` plus ``usage_example_c`` and ``controller_kind``. A linear plant is a condensed QP (``linear_qp``). A nonlinear plant, including differential drive, is a linearized real-time iteration (``linearized_rti``): Jacobians at the current state, then projected-gradient steps. That is not a full nonlinear SQP. Compile with ``gcc -O3 -std=c99 -Wall -Wextra -fPIC -shared -lm``. Each control step is ``mpc_compute_u0`` then, after the plant measurement, ``mpc_observer_update``. Neither function allocates on the heap. ``MPC_NX``, ``MPC_NU``, ``MPC_ND``, and ``MPC_HORIZON`` in the header are the array sizes. Same payment as autotune_template: X-API-Key spends 2 autotune credits, and a missing key or an empty pack is HTTP 402 for x402.
autotune_template1 of 1 parameter describedTune differential_drive, battery_peak_shaving, or root_zone_irrigation. This is POST /v1/templates/autotune at $2.00. The name is differential_drive, battery_peak_shaving, or root_zone_irrigation. The reply returns tuned Q, R, Rd, and Qf. Send name and max_trials. Send header X-API-Key when a $29 pack still has autotune credits. One call spends 2 of those credits. With no key, or when the credits are gone, the call is HTTP 402 and an agent pays with x402. The reply returns the tuned weights for that template. Payment uses the same gate as the HTTP route.
Tested by ConnectATD. Spotted a mistake? Tell us.