Summary

Check `ml_ready` before trusting the ML tools — they answer either way.

The neural solver-selection tools return a flag saying whether the model files were actually found, and the quick-start walks you through verifying it. That is a good design: the failure is visible rather than silent. The other thing to note is the engine split — the external `rust-optimizer` is a separate cargo install, and the README's own comparison table puts it well ahead of the internal engine on both speed and route length, so it is worth installing before drawing conclusions about quality.

What it is

A Rust route-optimisation engine with an MCP server built in. The pipeline runs from raw map data to an optimised route: extract a road network from Overture, OSM or PMTiles, compile it into a compact binary format, then solve either the Chinese Postman Problem — covering every street — or a Vehicle Routing Problem across a set of stops. Several solvers are available, and ML models can pick between them.

What you get

Pipeline tools including `extract_overture` for road data, `compile` for the binary map, and `optimize` for a CPP solve
`v2rmp_rust_optimizer` uses the external optimiser, which the README reports as faster and producing shorter routes than the internal engine
`vrp_solve` for vehicle routing, with greedy, Clarke-Wright, sweep, two-opt, OR-Tools and neural solvers
`partition` splits a graph into zones; `haversine_distance` computes point-to-point distance; `list_r2_bucket` lists stored artefacts
Three ML tools — `predict_solver`, `predict_quality` and `tune_hyperparams` — which report an `ml_ready` flag so you can tell whether the models actually loaded
The repository's own regression harness exercises 20 tools and reports 18 passing with 2 skipped; the two skipped ones, `pipeline` and `get_valhalla_matrix`, need internet access
A CLI covering the same ground, plus `serve` to start the MCP server

Requirements

Install with `cargo install v2rmp`, then point your client at the `rmpca` command with the `serve` argument. For better CPP results also `cargo install rust-optimizer`. The ML tools need a build with the `ml` feature and the model files present — `predict_solver`, `predict_quality` and `tune_hyperparams` report `ml_ready` as `False` when the weights are missing, which is the signal to train or copy them. No account and no key; the server speaks JSON-RPC over stdin.

Setup effort

One command — cargo install v2rmp