[BUG] Open WebUI + HA-MCP: Model fails to trigger tools, defaults to Knowledge Base/RAG

[steveof2620]

💬 What Went Wrong?

cannot get the openwebui model to work with or communicate with the home assistant mcp add-on / app.

Setup Details:

• Host: Open WebUI (Docker)
• HA Add-on: Home Assistant MCP Server
• Models Tested: qwen2.5, llama3-groq-tool-use
• Status: Connection is verified (200 OK in logs), but the model fails to execute tool calls.

The Problem: > Even with "Native" function calling enabled and tool security policies disabled, the LLM refuses to use the Home Assistant tools. It instead hallucinates that it needs to search a "Knowledge Base" or claims it "lacks the capability" to approve requests. It seems unable to recognize the MCP schema as an actionable tool set.

Steps Taken: > Summary of Troubleshooting Steps Taken

1. Infrastructure & Connectivity Verification
   Successful Connection: Confirmed the ha-mcp Add-on is running and successfully communicating with Open WebUI, verified by 200 OK status codes in the add-on logs.

Environment Tweaks: * Updated WEBUI_SECRET_KEY for secure communication.

Adjusted MCP_INITIALIZE_TIMEOUT to 60s to prevent session termination.

Tool Search: Tested both ENABLE_TOOL_SEARCH = true and false to see if the model could correctly process the entity list.

2. Open WebUI Configuration
   Native Mode: Enabled "Native" function calling in Chat Controls, as standard prompt-based injection was unsuccessful.

Injection Verification: Verified the presence (and occasional absence) of the "Home Assistant" badge/chip above the input box, confirming whether the tool was successfully injected into the chat context.

Security/Permissions: Confirmed "Enable Tool Security Policies" is set to off in the add-on settings to prevent "human-in-the-loop" approval blocking.

3. Model & Reasoning Challenges

Model Testing: Tested with multiple models, including qwen2.5:latest, llama3:latest, and llama3-groq-tool-use:latest.

Hallucination Issues: Encountered a consistent "Agentic Conflict" where models prioritize built-in "Knowledge Base" (RAG) tools over the Home Assistant MCP tools, even when explicitly instructed not to.

Schema Parsing: The models consistently respond that they "don't have the capability" or "cannot find information in knowledge bases," suggesting they are struggling to parse the JSON schema provided by the MCP server or are misinterpreting the agentic flow.

Reset Attempts: Performed multiple hard refreshes (Ctrl+F5) and initiated fresh chat sessions for every configuration change to rule out state caching.

Request: Has anyone successfully implemented MCP with these local models? Are there specific system prompt requirements to prevent the model from defaulting to internal RAG tools?

📋 AI-Generated Content / Logs / Config

Seen in the add-on logs
2026-06-18 09:30:41 INFO:     192.168.1.200:61935 - "POST /private_coLh71CLJ5TBF7kMIrYl6w HTTP/1.1" 200 OK
INFO:mcp.server.lowlevel.server:Processing request of type ListToolsRequest
DEBUG:mcp.server.streamable_http:Terminating session: None

🖥️ Client Application

Open WebUI

📦 Installation Method

Home Assistant Add-on

💻 Operating System

macOS

OS Version

macOS 26.5.1

📖 Setup Guide Followed

https://homeassistant-ai.github.io/ha-mcp/setup/

🔄 Exact Steps You Followed

All steps listed at
https://homeassistant-ai.github.io/ha-mcp/setup/

🐍 Python Version

No response

📦 ha-mcp Version

No response

🏠 Home Assistant Version

2026.6.3

💡 Additional Context

Have tried everything I can think of, just cannot get this to work with openWebUI. It does work with Claude Desktop though.

[github-actions[bot]]

✅ Thanks for the report — this issue has enough information and has been added to the backlog.

🤖 Automated triage by Issue Bot

[kingpanther13]

Have you tried increasing context window size in your openwebui? Also Enabling tool search on ha-mcp should help it work since that shrinks the idle context by a lot. I literally just closed an issue with this yesterday #1651 @MrChunks he is using a similar setup as you. Just refreshing after changing addon settings won't be enough, you probably need to disconnect and reconnect the MCP server completely. These small models are very finnicky to get working properly. You could also try using a newer model of Qwen too, when I experimented with this a few months ago I think I got it working with 3 but not 2.5.

[steveof2620]

Have you tried increasing context window size in your openwebui?
This is something I'll have to have a look at. Not sure how to do that TBO.

Also Enabling tool search on ha-mcp
I have tried this both enabled and disabled, doesn't seem to make a difference.

could also try using a newer model of Qwen too
Haven't tried Qwen 3, but I have tried with other models (for example llama3-groq-tool-use:latest). When I ask about home assistant I just get generic responses about how to install it and so on.

Just will not see the MCP server even though it has been setup multiple times and all looks green in openWebUI's integrations page.

Thanks for responding.

[kingpanther13]

Yeah I've had them aggressively deny that the mcp server exists before lol. The context window should be adjustable in your settings somewhere. Definitely leave toolsearch enabled either way for these small models.

[kingpanther13]

I'm going to convert this to a discussion, I don't think there's anything that we can really change on our side to help with this, this is all client behavior unfortunately... However, if I'm wrong and you find something that we can change on our side please reopen the issue or make a new one.

[Dulcow]

I'm trying pretty much the same, running a local model on my RTX 4060 Ti 16GB to debug, interact and configure Home Assistant. MCP endpoint seems to be up and accessible but all models fail to call/interact with it properly.

Any luck @steveof2620?