A new custom integration called Nexus MCP showed up on the Home Assistant community forum on October 15. It registers as a conversation agent and lets an AI model look up entities, call services and create automations. We have not run it ourselves, so everything here comes from the developer’s posts and the GitHub repository.
What it does
The developer lists four built-in tools: entity search, state lookup, service calls, and automation creation. Created automations are written to automations.yaml and hot-reloaded. It works in the Assist pipeline (web, mobile app, ESPHome voice satellites) and now has its own dashboard chat card with a model switcher and image paste.
Engines are Groq (the preconfigured default), Google Gemini, Qwen Cloud, and Ollama or a custom OpenAI-style endpoint. The developer cites free daily limits of 14,400 requests on Groq and 500 on Gemini. The README says Groq, Gemini and Qwen are verified in production, while Ollama and custom endpoints are experimental pending community testing.
The safety switch
By default, Autonomous Mode is off. The agent explains the planned triggers, conditions, actions and entities, then asks before writing anything. With it on, it creates automations immediately. We would leave it off, because a language model writing to your live config deserves a human glance first.
A fast-moving project
The releases page shows versions 1.3.0 through 1.4.0 landing around October 15, with 1.4.0 adding ten quota and rate-limit sensors. The developer says an official HACS default-store request is still pending, so for now it installs as a custom repository.
One forum user with about 3,500 entities reported a series of early hiccups: replies in the wrong language, a malformed JSON error, an entity-count problem, a dashboard card that would not load. The developer shipped fixes within hours. That is responsive, but it also tells you this is early software.
What owners are asking
- Will an update break my setup? It did for that forum user. Version 1.4.0 renamed entities, and an existing Assist pipeline pointed at the old one failed. The developer’s fix was to re-select the Nexus MCP agent under Settings, Voice Assistants, Pipelines, then remove the orphaned entity. A hard browser refresh was also needed for the dashboard card.
- Is it local? Only with Ollama. The Groq, Gemini and Qwen options are cloud services, so your prompts and the entity data the agent retrieves go to those providers. Home Assistant’s own docs make the general point that cloud services stop working if the company changes terms or goes away. We found no statement on pages we read about how long free tiers will last.
- Does it work with my Home Assistant version? The README lists a 2024.1.0 minimum and testing through 2026.9.4 and later. It does not mention 2026.10, so if you are weighing that release, read our earlier post on the reported 2026.10 rollback first.
What we would do
- Take a full Home Assistant backup before installing, since it edits automations.yaml.
- Start with Autonomous Mode off and read every proposed automation.
- Treat free-tier quotas as a budget. The new sensors show remaining daily requests, which is useful.
- Do not put it on a system that unlocks doors or controls heating you depend on until it has some history behind it.
- Wait for the HACS default-store listing if you prefer a vetted install path.
It is an interesting way to get natural-language automation building without a paid subscription. Try it on a test instance first, and keep your working automations backed up.