github mostlygeek/llama-swap v252

3 hours ago

I've been very busy over the past few weeks and haven't had much time for llama-swap. First, thank you for all the contributors in this release:

  • @sousekd for adding on_startup.profile. No more manual switching every time
  • @bernardogv for adding the -validate flag so configurations can be checked in one step
  • @chrispaulm for improving context limit compat in v1/models
  • @cphlipot for cmd/wol-proxy improvements
  • @ShinkoLab for improving IME (input method handler) support in the Playground
  • @WhyNotHugo for their keen eye catching a typo fix.

The key feature of is a Help agent in the Playground:

llama-swap-help-agent.mp4

I've finally accepted that writing documentation will never be prioritized over interesting engineering things. At the same time the complexity of the configuration can no longer be ignored. Having only config.example.yaml was fine in the beginning when there was only the models configuration but now there are dozens of knobs to configure concurrency, swapping, routing, remote peers, etc.

So now llama-swap has a built in Help agent. It's been tuned so a small agentic model can search and generate accurate answers to questions. It was fun to build:

  • an evaluation set so a frontier model can tune the performance of the a specialized agentic system. The baseline model, gemma-4-12B Q4_K_XL passes at 100%. It's a prototype and my first time using an eval set and an "agent in a loop" to optimize a whole system.
  • a core implementation of stateless MCP 2026-07-28 for doc search. Foundation for more llama-swap to be a stateless MCP proxy. No support for the stateful protocol stuff because I hate the design and think it should be deleted.
  • new UI elements for representing reasoning, tool calls, etc.
  • new docs/kb knowledge base that is easy to contribute to.

Changelog

Don't miss a new llama-swap release

NewReleases is sending notifications on new releases.