Home lab · Open WebUI and n8n

Already run a home lab? Plug chanio in.

If Open WebUI or n8n already runs in Docker on your Mac, it can read your chanio notes. chanio itself stays outside the containers, on the Mac, and a small bridge you start in Terminal lets them call it over HTTP, with a key you pick.

Why chanio isn’t in a container

Docker on a Mac runs its containers inside a small Linux machine, and that machine can’t reach the Mac’s GPU. A model in a container runs on the processor alone. Ollama installed as a normal Mac app uses the GPU, so chanio and Ollama run on the Mac itself, and your containers talk to them over the network.

The same goes for Open WebUI’s own chats: point its Ollama connection at http://host.docker.internal:11434 and it uses the Ollama on your Mac instead of a slower one inside Docker.

What your home lab can reach

chanio’s MCP server, the same one chanio mcp sets up for Claude Code or Cursor. It has seven tools, and six of them only read. It talks over stdio, the way a desktop app starts it, so for Open WebUI and n8n you put a bridge in front of it: mcpo, a small open-source program from the Open WebUI project that turns each tool into an ordinary web address.

The door is different. The page chanio door opens answers only itself: it checks that the request came to 127.0.0.1 on your Mac, and every call carries a key that changes each time the door opens. A container can’t call it, on purpose. Use the bridge.

  1. Step 1 · on the Mac, in Terminal

    Start the bridge

    You need uv, the Python tool runner from Astral (brew install uv if you use Homebrew). First make a long key for the bridge and keep it in this Terminal window:

    export CHANIO_BRIDGE_KEY=$(openssl rand -hex 24); echo $CHANIO_BRIDGE_KEY

    Then start the bridge in the same window. It runs until you close the window or press Control-C.

    uvx --with 'mcp<2' --with 'fastapi<0.120' mcpo --host 127.0.0.1 --port 8000 --api-key "$CHANIO_BRIDGE_KEY" -- chanio mcp --serve

    It listens on 127.0.0.1, so only this Mac can reach it, and every call must carry the key. The two pins (mcp<2 and fastapi<0.120) are there because the newest mcpo, 0.0.20, does not start with the newest releases of those two libraries as of November 2, 2026. With the pins it starts and answers.

    Check it from a second Terminal window (paste the key in place of the variable there, or export it again):

    curl -s -X POST http://127.0.0.1:8000/search_notes -H "Authorization: Bearer $CHANIO_BRIDGE_KEY" -H 'content-type: application/json' -d '{"query":"pricing"}'

    The answer is the matching passages, as one string. Every tool and its input is listed at http://127.0.0.1:8000/docs.

  2. Step 2a · Open WebUI

    Add chanio as a tool server

    Add it as your own tool server, not an admin one. Open WebUI calls your own tool servers from your browser, so localhost is your Mac, and nothing about Docker’s network matters.

    • If you are the admin, first turn on Direct Integrations: Settings, Admin, Integrations, under User Connections. It is off until you do.
    • In your own Settings, open Integrations and add a tool server. URL: http://localhost:8000. Auth: Bearer, with your key.
    • In a chat, press the Integrations button beside the +, choose Tools, and turn chanio on.

    This works when you open Open WebUI at a localhost address on the same Mac. From another device on your network, the browser would look for the bridge on that device instead.

  3. Step 2b · n8n

    Call a tool from an HTTP Request node

    n8n makes the call from inside its container, so the address is host.docker.internal, the name Docker Desktop gives your Mac. In an HTTP Request node, choose Import cURL and paste:

    curl -X POST http://host.docker.internal:8000/search_notes -H 'content-type: application/json' -d '{"query":"pricing"}'
    • Authentication: Generic Credential Type, then Bearer Auth, with your key as the token. The key then lives in n8n’s credentials, not in the workflow.
    • To ask from earlier in the workflow, change the body to an expression such as {"query": "{{ $json.question }}"}.
    • Any other tool is the same call with its name at the end of the address and its input as the body, from the list below.

    If n8n says the connection was refused, the container can’t reach a bridge that listens only on 127.0.0.1. We have not checked that case on a Mac yet. Start the bridge with --host 0.0.0.0 instead: then other devices on your network can reach it too, and the key is what keeps them out.

The seven tools

Each is a POST to the bridge, at the tool’s name, with a JSON body.

search_notes {"query": "…"}

The passages from your notes, saved pastes and any folder you let chanio read that match the words, with the file each came from. Passages that look private are left out, and the reply says how many.

topic_card {"topic": "…"}

A context card on one topic: only non-private passages and its open threads, with a SHA-256 fingerprint of the card.

read_brief {}

The latest morning summary, private lines left out.

memory_pack {}

Your declared aims and the last week of summaries, compiled on the Mac.

reaim_status {}

What you said you are aiming at, next to what you actually wrote about in the last 14 days.

meter_today {}

What was answered where today, and anything it cost.

hold_question {"question": "…"}

The one tool that writes. It adds the question to your kept list, marked as added by an AI app, and it waits there until you confirm it with chanio keep or in the door.

When a call fails

From chanio 0.16.1, a failed call says what went wrong, the fix, and the command to run. Through the bridge it arrives as an error answer with those lines inside it. An empty question to hold_question, for example, comes back with question is required, then Fix: Pass the question as {"question": "…"}. and Run: chanio mcp. Earlier versions give the first line only. chanio errors --json lists every error and its fix.

What leaves the Mac

The bridge hands each tool’s answer to the app that called it, with passages that look private already left out, exactly as chanio mcp does for Claude Code. What the app does next is up to the app. If Open WebUI answers with the Ollama on your Mac, nothing leaves it. If it answers with a model in the cloud, the passages go to that model with your question. Stop the bridge and nothing can call chanio at all.

Not set up yet? Open your mirror in this tab, or one line puts chanio on your Mac. Every command on this page was run against chanio 0.16.1 and mcpo 0.0.20 on November 2, 2026.