Live in your browser, solving now

Loading the solver…

Routing that runs in your account.

Everything above solved in your browser, the same metaheuristic you would run in production. In your VPC it runs on your data, with no per-call fee. Sold on AWS Marketplace as a container you deploy in minutes.

See it on Marketplace How it works

180 stops solved live above, in your browser tab
0.7% typical gap to the best plan we can find
0bytes of data that ever leaves your account
48MB container image, cold start under a second

The cost is in the matrix.

A hosted routing API meters the distance matrix, stops² elements, priced on every call. A 500-stop plan is a 250,000-element matrix. Re-plan through the day and you pay for it again, and again. Arclane computes the matrix once, in your account, and never charges you to re-plan.

Put in your own numbers. The annual figure recomputes as you type.

Loading calculator…

Ingest

Stops, windows, demand, skills, vehicles, depots, shifts.

Matrix

Built once, in your VPC, from your coordinates or road times.

Search

Construction, then local search and ruin-and-recreate.

Validate

Capacity, windows, skills, duty time. Nothing illegal ships.

Commit

Run sheets, reason codes, and a locked-stop replan.

Your rules, not a fixed menu.

The solver enforces the constraints your dispatchers already work around by hand. Because it runs in your account, the constraint model is yours to extend, not a dropdown of the vendor's choosing.

See the constraint model

  • Promised arrival windows, hard and soft
  • Vehicle capacity, weight and volume
  • Driver shift and duty-time limits
  • Skills and certifications per stop
  • Multiple depots and mixed fleets
  • Locked, completed stops on a live replan

When a stop can't be served, it says why.

No silent drops. Every unassigned stop carries a machine-readable reason code and a plain sentence, the same JSON your systems consume and your dispatcher reads.

Read the explainability model

{
  "stop": 142,
  "status": "unassigned",
  "reason_code": "WINDOW_INFEASIBLE",
  "message": "No vehicle can arrive before this stop's window closes.",
  "window_start": 545,
  "window_end": 605
}

It slots in where your data already lives.

Arclane is a container and an API. It reads the systems you run and writes back a plan, without a new source of truth and without shipping a byte outside your account.

INGEST

TMS & OMS

Pull stops and orders from your transport or order system by CSV, JSON, or API.

MATRIX

Your road network

Use straight-line distance or feed your own OSRM or Valhalla travel times.

DISPATCH

Driver apps

Push run sheets and live re-plans to the apps your drivers already carry.

TELEMETRY

Your observability

Solve metrics as plain Prometheus. No vendor dashboard to log into.

IDENTITY

Your IAM

Runs under your roles and network policy. No outbound calls to allow-list.

SCALE

Your cluster

One pod or many. Plans run in parallel across replicas you control.

We publish the gap, including our losses.

Each bar is the price of stopping early: a quick solve against a thorough solve on the same instance. Hover any bar for the full row.

Loading chart…

Full table, methodology, hardware, and the reproduction command →

Three lines to a running solver.

helm repo add arclane https://charts.zecehold.com
helm install arclane arclane/arclane \
  --set solver.replicas=3

Multi-arch images. Runs on Graviton. Deployment guide →

Container image
48MB
Cold start
0.9s
Architectures
amd64 · arm64
Data egress
0bytes

Why not just use OR-Tools?

For a batch job on one machine it is excellent. What it does not give you is a solver that runs in a browser tab, a packaged container with a Helm chart and reason codes, or a support contract. Arclane is the productised, self-hostable version of what you would otherwise assemble and maintain yourself.

Is it actually fast?

On 8 vCPU we reach a low-single-digit gap on a 1,000-stop instance in about ten seconds. Latency depends on your hardware, not our queue. The benchmarks show every number.

Is this "AI"?

No, and we will not pretend otherwise. It is a metaheuristic: construction plus local search plus a ruin-and-recreate kick. In a market that calls constraint programming "AI," saying what it is precisely is the differentiator.

How does procurement work?

Through AWS Marketplace, with private offers, and spend counts against your AWS EDP drawdown. See pricing.

Bring last month's run sheet. Get the number in seconds.

Upload a CSV. See vehicles used, distance, duty hours, and the cost delta versus what you actually ran. No account, nothing uploaded.

Benchmark your routes