Skip to content

Model Farm 0.8.0

Latest

Choose a tag to compare

@acsdd acsdd released this 16 Nov 08:29
· 13 commits to main since this release

Model Farm 0.8.0

Ollama is not the only local runtime, and it is not always installed. This
release finds the runtimes that are already on the machine, lets you pick the
one you want to use, and installs Ollama for you when you want a fresh
version.

What's changed

Local runtimes: find what you already have

  • A new Local runtimes card in the Model Farm scans the machine for
    Ollama, LM Studio, llama.cpp, Jan, GPT4All, KoboldCpp, LocalAI, vLLM,
    llamafile and text-generation-webui
    , and reports each one as installed,
    running or missing.
  • A runtime is found three ways: a server answering on its well-known port
    (11434, 1234, 8080, 1337, 4891, 5001, 8000, 5000), its command line on
    PATH, and the program in its install folder.
  • A leftover data folder is not an install. An app you uninstalled but
    whose settings remain is reported as missing, so the panel never sends you
    to a program that is no longer there.
  • The server banner names the runtime that holds a port. When two runtimes
    share one — llama.cpp, LocalAI and llamafile all want 8080 — the one that
    identifies itself wins and the others say who holds it.
  • Versions are read from the CLI, including the ones a runtime prints on
    stderr (llama-server does), and from the banner a runtime prints when it
    has no semantic version (LM Studio answers CLI commit: 43d4b26).

Use the runtime you want

  • use this makes a runtime the one the Model Farm installs into. The
    choice is stored in ~/.toolbox-mcp/settings.json and survives a restart.
  • The Model Farm follows the choice. Ollama pulls a GGUF straight from the
    Hub; a runtime that takes files instead of pulls (LM Studio, Jan, GPT4All,
    KoboldCpp, text-generation-webui) gets the weights downloaded into its own
    models folder; a runtime that serves one model per process (llama.cpp,
    llamafile, vLLM) is started with the model it was given.
  • The install action now reads Install in <runtime>, so it always
    names the runtime that will actually receive the model.

Install or update Ollama

  • Install / update Ollama downloads the official installer and runs it
    silently — per user, so no administrator rights are needed — then starts
    the server and confirms the version.
  • It always installs from the official source, so you get a fresh version
    rather than whatever was on the machine.
  • Runtimes with their own installer get a get it link to their homepage
    instead of a button that would fail.

Five new MCP tools

  • runtimes_detect — every runtime on the machine, with its state, version,
    path and the models it serves.
  • runtimes_install — install a runtime from its official installer.
  • runtimes_use — pick the runtime the Model Farm installs into.
  • runtimes_models — the models a runtime is serving right now.
  • runtimes_download — download a GGUF into a runtime's models folder.

The tool count goes from 24 to 29.

Notes

  • Detection never launches a desktop app. A program found in its install
    folder is reported; only a command on PATH is executed for the version
    probe.
  • The runtime choice is the only new setting, and it lives in the same
    owner-only settings file as the rest.
  • Nothing about this release changes how the Gmail App Password is stored —
    it still never leaves ~/.toolbox-mcp/settings.json.