Xval Engine

A small model on your Mac gives Claude only the skills each task needs; full tool routing arrives in v0.4.1. Opt in, measured on your own sessions.

Every skill, server and tool you install is described to Claude on every turn, whether the task needs it or not. Xval Engine is a small model that runs on your Mac and, before an agent starts, picks the ones the task actually needs.

Turn it on

It’s off by default. Turn it on in Settings › Xval Engine. Turn it off the same way. Reset learned priors clears what it has learned from your usage.

What it does

  • It reads the task’s first prompt and routes once, when an agent starts. What Claude sees then stays steady for the rest of the session.
  • The ten skills most likely to matter get their full descriptions. Every other skill stays listed by name, one read away.
  • It fails open. If Xval is missing or slow, the agent starts with its normal, full setup.
  • A resumed session keeps the route it started with.

Your MCP servers

In v0.4 Xval Engine routes your skills. Your MCP servers load exactly as before, and your settings files stay as they are.

  • Full tool routing, where Xval Engine also hands Claude only the MCP tools a task needs, arrives in v0.4.1.
  • When it does, your own permission rules and any managed policy will apply to every tool exactly as they do today.
  • If anything isn’t ready, the agent launches the way it always did, and the reason is recorded in the local log.

See your own numbers

The Settings › Xval Engine tab shows what it did on your Mac: sessions routed, first-call tokens saved and fallbacks. These come from a local log of counts and timings, never from task text.

What we measured

With full tool routing, on a 32-task benchmark of hard, multi-step coding tasks, cost per task was 15–20% lower than stock Claude Code, with the same success rate within what 64 runs can show. Very long sessions save less: past about 600k tokens the conversation itself is most of the bill, and the saving falls to about 0–3%. The method, the benchmark and the caveats are on the Xval Engine page.

Private by design

The model runs on your Mac and learns from counts of what you use, never from copies of your conversations. Nothing is sent to us. See Privacy.

Still stuck? Write to us at vaiditya@xagents.in. We read everything.

Last updated October 11, 2026.