help with ollama on intel

[foolsteakdave]

Has anyone ever managed to actually get goose working with ollama? If so, how exactly step by step?

I need guidance on what can realistically be expected, and what cannot.

Yes, I plowed through the documentation, the issues, and the pull requests, I tried every possible combination of configurations, and suggested steps. Yet it seems to be impossible to get goose actually working with ollama. - Neither on macOS (intel based), nor on ubuntu server 24.04 lts. And from browsing the issues and pull requests I don't get the feeling this will change any time soon.

So, primarily I want to understand if there is value in hoping for improvement for this specific use case.

Secretly I hope it's my fault I can't get it to work, so I could fix it, but the issues and pull requests give me the impression I am not the only one.

But to be clear, I am not asking for development effort to serve my use case, but rather crystal clear understanding of what can be realistically be expected, and what cannot.

Lastly should my problem be a case of "rtfm", please drop the link(to the manual)! I'd really appreciate, if I was to blame, because then it would be within my powers to fix it.

Thanks in advance!

(Background: goose seems never to be able to establish a working connection to ollama.

Error: Request failed: error sending request for url (http://localhost:11434/v1/chat/completions)
Interrupted before the model replied and removed the last message.
  error: The error above was an exception we were not able to handle.
These errors are often related to connection or authentication
We've removed the conversation up to the most recent user message
- depending on the error you may be able to continue

And also 404's on macOS.)

[zanesq]

@michaelneale might have some insight on this

[michaelneale]

@foolsteakdave how much ram is on your intel mac - is that the same one you are trying ubuntu with? (intel macs I don't think would have enough capacity for tool calling practically, maybve serving some of the smaller models but would need a bit more work before it could work with tool calling in my experience) - I don't think on an intel mac ollama would work for this (But am curious what your specs are as I could be wrong). A few of us tried on differnet hardware and can't quite reproduce this (with various models)

I would recommend gptoss:20B on ollama (as a starting point) but I am pretty sure that wouldn't work in this case with intel as there isn't enough GPU horsepower (there are some smaller qwen ones which may work, but very limited below that size). There is/was a one click setup for ollama which set that up, but it is hidden as it really requires some GPU or apple silicon with unified memory in 64g range to start.

[foolsteakdave]

Thank you @zanesq for routing my inquiry, and thank you @michaelneale for taking the time to answer it.

First some answers to your questions:
1.

how much ram is on your intel mac

--> 16GB
2.

is that the same one you are trying ubuntu with?

--> No
3.

But am curious what your specs are

--> Intel MacBook Pro 2016, Dual Core i7 3,3 GHz, 16 GB RAM, macOS Sonoma (via oclp) running ollama & goose-cli installed via homebrew
--> Shuttle DS77U5 as home server with an 2c/4t intel i5-7200u 3,1 GHz, 16 GB RAM, ubuntu server 24.04 lts running ollama as dockerised service available to the local network. I also tried, and failed to get goose-cli and ollama working baremetal on that machine before I deployed it as my new home server.

Is this sufficient information for you to build an opinion, or which further details do you need?

I am running qwen2.5-coder:latest as my go to model for coding.

Interestingly zed (the code editor) is able to make use of the ollama instance on my home server, as well as on the MacBook directly. It is by no means fast, but it works. Slowly, but it works, which is fine by me.

In contrast it seems to me that it is almost impossible to get goose working with ollama locally. - Which to my understanding is one of its key selling points, isn't it?

So, do I understand you correctly, that one is not to expect to be able achieve a working setup with goose and ollama unless one uses

some GPU or apple silicon with unified memory in 64g range to start.

, right?

If so, shouldn't that be incorporated prominently in the documentation? I assume that would also avoid the numerous (github) issues centering around that very topic.

... or is there hope that people with lesser hardware will have the chance to enjoy the "woolly pegasus" - as in goose & llama united - someday in the future?

Did I miss anything?

Please help me, and everyone else with the same problem to set realistic expectations.
If my use case is not supported by goose, that's fine. Yet, I need to know explicitly, because until now I was under the impression it should have worked.

PS. latest iteration as of today:

[REDACTED]@MacBook-Pro [REDACTED] % goose 
starting session | provider: ollama model: qwen2.5-coder
    logging to /Users/[REDACTED]/.local/share/goose/sessions/20250923_122525.jsonl
    working directory: /Users/[REDACTED]/Desktop/Projekte/[REDACTED]

Goose is running! Enter your instructions, or try asking what goose can do.

Context: ○○○○○○○○○○ 0% (0/128000 tokens)
( O)> Please run "npm test" and tell me how we can address the issues mentioned in the the error message! Thanks!
◐  Updating understanding units...                                                                                                                                                                                                          
Error: Request failed: error sending request for url (http://[REDACTED].internal:11434/v1/chat/completions)
Interrupted before the model replied and removed the last message.
  error: The error above was an exception we were not able to handle.
These errors are often related to connection or authentication
We've removed the conversation up to the most recent user message
- depending on the error you may be able to continue


Context: ○○○○○○○○○○ 0% (0/128000 tokens)
( O)> Press Enter to send, Ctrl-J for new line

The .jsonl is empty.

q.e.d.

[michaelneale]

@foolsteakdave I'm not sure if the same models will work with goose - can you tell me more about the zed setup - as that would be very close. It is interesting that it seems to be reaching to the ollama server.
Goose will work by default with gptoss:20b, but not just any model, lots of people do use it, but not just any model as it uses tool calling more heavily than others.

... or is there hope that people with lesser hardware will have the chance to enjoy the "woolly pegasus" - as in goose & llama united - someday in the future?

don't understand?

edit: oh llama + goose ... haha yes!

[michaelneale]

my config can talk to qwen2.5-coder, but it's tool calling is not quite right:

my config.yaml:

GOOSE_MODE: auto
GOOSE_PROVIDER: ollama
GOOSE_MODEL: qwen2.5-coder

security_enabled: true

OLLAMA_HOST: localhost

extensions:
...

trying the default one and it does work with basics (not sure what -coder ads - but at least the default size seems to not be as finely tuned for tool calling - if you put things in non auto mode, or chat mode - it should work however)

I like zed a lot - but haven't been able to make it work with ollama (it does same thing goose does - ie not really a tool call), would love to know your setup (I know it tries to use openrouter as well). I also like how it automatically turns off tools if models don't support it.

Hopefully can make them talk

[pjv]

Not sure if this is still an issue, but I hadn't seen this discussion and had some bumps getting goose working with ollama which I have running on a different host. The error messages I was getting were kind of misleading and what finally worked for me was including the port in the OLLAMA_HOST variable.

i.e.:

OLLAMA_HOST: http://pjv-mba.my_lan:11434