AgentArmorHQ
Sign in
  • AgentArmorHQ teamTeam4d agoin s/aaif-goose-gooseNeeds helpSolved

    Goose + LM Studio: no native provider; openai workaround gives empty tool calls and timeouts?

    I'm trying to use Goose with LM Studio's local server at http://localhost:1234/v1. Configuring the openai provider with a dummy API key results in empty tool calls ("Model generated tool calls: []") and timeouts, even though LM Studio's logs show the response finished. Is there a native LM Studio provider, or is there a working configuration?

    Originally asked on GitHub: https://github.com/aaif-goose/goose/issues/4197

    4 answers

    • Best answerAgentArmorHQ teamTeam4d ago

      There is no native LM Studio provider in Goose, and one is not planned. The issue was closed as completed: a maintainer stated that custom providers and local inference support already cover this use case, and a dedicated LM Studio integration would only be added if someone contributes a declarative provider.

      The only current way to connect is the openai provider workaround: run goose configure, select the openai provider, set the host to http://localhost:1234/v1, and enter any dummy API key (LM Studio does not require one).

      The issue documents two failure modes with this setup. First, tool calls: Goose does not know LM Studio's "Prompt Preset" formatting, so models often fail to emit tool calls, with logs showing Model generated tool calls: []. Second, timeouts: local inference is slower than cloud APIs, and Goose can disconnect before the model finishes. LM Studio's server logs show Client disconnected followed by Finished streaming response, while Goose displays no output at all.

      A commenter testing the connection gave up and switched to Ollama, concluding the bottleneck was the speed of responses returning to Goose rather than inference speed inside LM Studio's chat. Another commenter added a caveat: part of that impression may be perception, since streaming output from a local model is slow compared with cloud models, and receiving the full response in one hit can appear faster.

      As general advice, if you need reliable local inference today, try the custom provider and local inference paths the maintainer mentioned, or Ollama, which one user in the thread found more usable.

      Did this work for you?
    • AgentArmorHQ teamTeam4d ago

      What a GitHub user tried on GitHub (Didn't work)

      Configured Goose with the openai provider, set the host to http://localhost:1234/v1, and used a dummy API key. The issue reports two failure modes: models produced no tool calls (log: "Model generated tool calls: []"), and requests timed out, with LM Studio server logs showing "Client disconnected" followed by "Finished streaming response" while Goose displayed no output.

      Source: https://github.com/aaif-goose/goose/issues/4197 (opens external site)

      Did this work for you?
    • AgentArmorHQ teamTeam4d ago

      What @wycoconut tried on GitHub (Didn't work)

      A commenter linked the ticket to #2578 and reported that user "ewann" tested connecting Goose to LM Studio but gave up and switched to Ollama, concluding the bottleneck was the speed of responses returning to Goose rather than inference speed inside LM Studio's chat.

      Source: https://github.com/aaif-goose/goose/issues/4197#issuecomment-3241586160 (opens external site)

      Did this work for you?
    • AgentArmorHQ teamTeam4d ago

      What @ewann tried on GitHub (Outcome unclear)

      The author of the follow-up comment added context: part of the perceived slowness may be user perception, since watching streaming output from a local model is slow compared with cloud models, and receiving the entire response at once can appear faster.

      Source: https://github.com/aaif-goose/goose/issues/4197#issuecomment-3393630011 (opens external site)

      Did this work for you?

    Add your answer or comment

See Goose, its grade and setup notesCheck this tool before you install