Questions, answered from the code.
Including the ones whose answer is no. Every answer here is either a fact the test suite enforces or an admission — and each one names the page where you can check it.
Does the browser produce the same plan as the command-line tool?
Yes, byte for byte. The same order produces the same arrangement and the same SHA-256 content hash in the browser, in the CLI and in the Python package, and CI compares both engines on every commit, strategy by strategy. That is what makes a plan citable: a hash in an email names one arrangement that passed one set of thresholds.
What does the stability score actually measure?
Stability is 0.6 × the mean support ratio + 0.4 × a centre-of-mass score. Support is measured against the surfaces a case actually rests on, within a 2 mm contact tolerance — not by layer index, so a case suspended over a gap does not score as supported. It models geometry, mass and support; it does not model friction, wrapping, deck condition or transport dynamics.
Can an AI agent make this move a robot?
No, and not by any argument it can make. Authorizing motion needs a capability token minted inside the controlling Python process; no MCP tool, HTTP route or model output can mint one or ask for one. Two tests enforce it: one scans the tool registry for motion verbs, the other asserts no unregistered code path sends motion at all.
Is any of this certified?
No. Certification is granted to an assessed installation by an accredited body, never to a software package — so no software can certify itself, and any page describing a "certified connector" for this project is wrong. The CE conformity assessment and the risk assessment of a cell remain the integrator's and the operator's.
Has it run on a physical robot?
Not once. The Universal Robots adapter speaks the controller's dashboard and script interfaces and passes the adapter contract suite against an in-repo fake controller and against URSim in CI; its tier is COMMUNITY. FANUC and KUKA are PROFILE tier: cited reach and payload figures, no adapter. A VALIDATED tier needs a documented physical run, and nothing has one.
What is free, and what do I pay for?
The engine, the CLI and the MCP server are Apache-2.0 and unmetered — run them on your own machines and none of the limits on this site apply. What is paid is compute and signatures on our hardware: plan history, API keys, URScript and CSV exports from the API, larger orders, and signed verification attestations. Every paid line is an entitlement the code enforces.
Someone sent me a pallet plan. How do I know it is real?
Drop it on the plan checker. It recomputes the hash to catch an edit, and then — separately — re-runs the validator over the geometry and compares the result with the verdict the document claims. The digest is unkeyed, so anyone who can edit a plan can reseal it; only re-deriving the verdict catches that kind of forgery.
Can it plan an order that needs more than one pallet?
Yes: an order is packed onto as many pallets as it needs, each a separately sealed and verifiable plan. The split is greedy, not optimal, and every result publishes a lower bound on the pallet count so you can see how far from optimal it is — on one mixed 300-case reference order it uses six pallets against a bound of four, where the first version of the splitter used seven.
What can it export to a robot?
A URScript pose list and a placement CSV, both byte-identical to what the CLI writes. The pose list carries coordinates and the axis-angle rotation vector in the robot base frame — it is not a runnable program: no gripper I/O, no safety configuration, no controller preamble, and it authorizes nothing. The exporter refuses a plan whose hash or verdict does not hold up.
What do I actually hand a customer who asks me to prove the load was planned?
A proof pack: one archive with the sealed plan, the placement CSV, the URScript pose list, a signed statement about the plan, and a README telling them how to check every file offline with the open-source CLI. The bytes are deterministic, so two people can compare packs by hashing them. It is evidence of what was checked — not a certificate, and it authorizes nothing.
Can I run it entirely on my own infrastructure?
Yes. Install the package, plan from the CLI, and point your agent at the MCP server over stdio; nothing in that path calls this site. The hosted service is for the things that need somebody else's hardware or key — signed attestations above all, since the signing key never leaves the deployment.
Can the person I send a plan to check it without an account?
Yes, and that is the point of the signature. A plan file is checked in their browser at /plan; a signed attestation is checked at /verify, free and without signing in, against the public key this service publishes — or offline with the open-source CLI, against a key they pinned themselves. You pay to produce the statement; nobody pays to believe it.
What does this site store about me?
Nothing at all unless you sign in or send a pilot request: planning on the demo page happens in your browser. With an account, the site stores what the plan you are on says it stores — plan history, API key hashes, daily usage counters — each listed with where it lives and how long it is kept.
Who is behind it, and how far along is it?
One engineer, working in the open, pre-revenue, with no deployments and no customers to name — which is why every figure on this site is computed live from the engine on a named dataset rather than quoted from a case study. The current ask is a first 30-day pilot with a kill date, starting from a plan made on the buyer's own SKUs.
Every answer above is checkable on your own data: plan one of your real orders and compare it with what you ship today.
The engine is Apache-2.0 and unmetered on your own machines — what the hosted plans add is a signature only this service can make, state it keeps for you, and compute on its hardware.