All articles
Ollama & llama.cppFixes

Ollama “does not support tools”: Check Capabilities, Then Switch Model or Fix the Template (Tested)

"does not support tools" means the model lacks the tools capability. Check Capabilities with ollama show; the fix is a model with the Tools tag, or a tool-aware template for a GGUF you imported. Error and template fix tested on Ollama 0.40.15; real tool-call quality not tested.

Ollama does not support tools: a clay llama holding a wrench and a plug that do not match a clay toolbox, with a red clay cross
Also available in中文Español

Short answer: registry.ollama.ai/library/<model>:latest does not support tools means the model you picked does not declare the tools capability, so Ollama refuses any request that carries a tools array. Check with ollama show <model>: the Capabilities list has to include tools. The clean fix is to switch to a model that lists the Tools tag on ollama.com. If you imported your own GGUF, the cause is usually a Modelfile TEMPLATE that never mentions tools. I reproduced the error on Ollama 0.40.15 with a tiny test model and tested both the error and a template fix. I could not test real tool calling quality, because the test model file is only 1.19 MB (about 6 MB once loaded; stories260K.gguf on Hugging Face).

What the error says

Here is the error as I got it, from both Ollama’s native API and the OpenAI-compatible one. Both returned HTTP 400:

{"error":"registry.ollama.ai/library/tiny:latest does not support tools"}
{"error":{"message":"registry.ollama.ai/library/tiny:latest does not support tools","type":"invalid_request_error","param":null,"code":null}}

The request that triggered it was an ordinary chat request with a tools field added. The same model answered the same chat without tools with HTTP 200. So the model works; only the tool part is refused. Many agent tools and coding assistants send tools with every request, so they hit this error as soon as you point them at a model without the capability.

Step 1: check the model’s capabilities

ollama show <model>

For my test model the output ended like this:

  Capabilities
    completion

No tools line means any request with tools will fail. You can read the same list from the API with POST /api/show, in the capabilities field (mine returned ['completion']). For a model that works, the list includes tools.

Fix 1: use a model that has the tools capability

This is the right fix for almost everyone. Ollama’s model search has a tools category: ollama.com/search?c=tools. Pick a model from that list that fits your memory, pull it and point your tool at it. Ollama’s tool calling docs use qwen3 in their examples. For sizes and memory, see the best local LLMs and local LLMs for coding.

Fix 2: add a tool-aware template to a model you imported

If you imported a GGUF with your own Modelfile and its TEMPLATE has no tool handling, Ollama does not list the capability. I tested this. With this Modelfile the capability was missing:

FROM ./model.gguf

and with this one, which only mentions .Tools in the template, the Capabilities list gained tools and the same request no longer returned the error:

FROM ./model.gguf
TEMPLATE """{{- if .Tools }}Tools: {{ .Tools }}
{{ end }}{{- range .Messages }}{{ .Role }}: {{ .Content }}
{{ end }}assistant: """

After ollama create my-model -f Modelfile, ollama show my-model listed completion and tools.

Read the limits of this carefully:

  • I only tested that the error goes away. My test model is far too small to call a tool, so I cannot say anything about call quality.
  • Ollama’s Modelfile docs list .System, .Prompt and .Response as template variables and do not describe .Tools. What I saw is the behavior of version 0.40.15. Treat it as an observation, not a documented guarantee.
  • A model that was never trained for tool calling will not become good at it because its template mentions tools. Use the real template from the model’s authors, not a made-up one like mine.

What it is not

More messages are in the Ollama and llama.cpp error list.

FAQ

Can I turn tools off in my client instead?

If your client lets you disable tool or function calling, the request will not carry tools, and my test model answered such a request normally. Many agent tools need tools to work at all, so a model with the capability is the real fix.

Why does a model that supports tools elsewhere fail on Ollama?

From what I saw, Ollama reads the capability from the model’s own metadata and template. The same weights imported without a tool-aware template showed no tools capability in my test.

Does the OpenAI-compatible endpoint behave differently?

Not here. Both endpoints returned HTTP 400 with the same message text.

Sources and test notes

  • Ollama docs: Tool calling (docs.ollama.com/capabilities/tool-calling) and Modelfile reference (docs.ollama.com/modelfile).
  • Tested myself: Ollama 0.40.15 on Linux with the 1.19 MB stories260K.gguf test model: the error on /api/chat and /v1/chat/completions, ollama show, /api/show, and a template with {{ .Tools }}.
  • Not tested: real tool calls, any model from the library, Windows or macOS, client apps.

Related: Ollama GPU requirements by VRAM.

Leave a thought

Your email address will not be published.