[Local LLM 23] Check OpenClaw’s Ollama Endpoint and Tool Calls

Building a Local LLM · 23/30 · Connect with an Agent

Full series and learning sequence · Practice code ZIP

When chat works but tool calls fail, inspect the provider settings and API path. The OpenClaw Ollama provider documented here uses the native API. Copying the endpoint used by Hermes can therefore select the wrong path, even though both applications connect to Ollama.

With Part 22 installed, prepare the tool support model. It also checks the memory to load the model and context cache.

Ollama native API address

The baseUrl is specified as http://127.0.0.1:11434, and the API type is specified as ollama. /v1 is not appended. This is different from Hermes’ custom endpoint. If the tool call JSON outputs like a normal answer, look at these two values ​​first.

Invoke-RestMethod http://127.0.0.1:11434/api/tags
openclaw onboard --classic

Select Ollama → Local only to specify the local address of Windows. If your model is not listed, follow the instructions to manually select it with the correct tag. Automatic discovery can check not only files on disk, but also memory loading status and tool/context support conditions, so it does not always appear that it has just been downloaded.

Execution context and memory

The Ollama workbook recommends a minimum of 64K context for OpenClaw. We need memory to hold the KV cache along with the model weights. The 16K setting below is for checking connectivity for small requests. Do not use as a recommended specification for long agent jobs.

FROM qwen3:4b-instruct
PARAMETER num_ctx 16384
PARAMETER temperature 0.2

ollama create lab-qwen-16k -f .\Modelfile.agent
ollama run lab-qwen-16k "연결 확인이라고만 답해 줘."
ollama ps

Select this model during onboarding and start with a short query. If the context is filled with system instructions and tool definitions alone, you need to reduce the work or free up resources to run the larger context. Entering an unsupported maximum value in the settings will not increase the model’s limit.

Model discovery and tool invocation

openclaw models list --provider ollama
openclaw gateway status
openclaw dashboard

Once you’ve summarized your notes, test one read-only tool on a non-sensitive folder. Check your dashboard history for requests and results. The model’s description of “the file was read” does not determine actual execution.

Permissions are limited by the current version’s tool approval/sandbox settings. Simply writing “Read Only” as the prompt does not remove actual write access. When executing a command or sending it externally, the execution details can be reviewed.

Web search/fallback connection check

Web browsing can be automatically set up in Ollama’s launch openclaw path. This lab does not use web searches or external channels. We also check the fallback provider and auxiliary embedding addresses to see what functions are connected in addition to local inference.

openclaw configure --section web
openclaw gateway stop

If the configuration does not use the cloud function, also review the OLLAMA_NO_CLOUD setting in Part 29. However, this setting does not block external communication from other apps. After making changes, double-check that your local answers and tool calls still work.

reference material

• OpenClaw Ollama Provided by
• Provider Onboarding/Discovery Terms
• Ollama Provider Settings
• Ollama OpenClaw Integration

Documentation and model tags checked: October 18, 2026. Primary lab environment: Windows 11 x64.

---