Windows local-agent preflight checklist

One page of checks before you install a local model stack on Windows. No email needed. Nothing to sign up for. Tick a box only after you have looked.

The tested path is Ollama for the model, Open WebUI for the chat page, and n8n for workflows. A failed row means you are outside what was tested, or something else is already using a port the stack needs.

The install guide does not ask you to run these steps as a Windows administrator. The tested install left the user PATH unchanged and did not add a sign-in startup entry. You start each program from a script in the stack folder.

Windows version and admin accounts

Windows 10 or Windows 11. Open Settings, then System, then About, and read the edition. The scripts and the tested steps are not written for Mac or Linux.
You will use Windows PowerShell, the one the start scripts call. Nothing in the tested install registers the stack to start at login.
As soon as http://127.0.0.1:8080 answers, create the first Open WebUI account yourself. That account becomes the admin. Until it exists, any program on this PC can claim it. "onboarding":true means no admin exists yet.
After you are the Open WebUI admin, turn new signups off under Admin, Settings, General. The tested config still had "enable_signup":true when the install stopped, so this click is yours.
As soon as http://127.0.0.1:5678 answers, set the n8n owner account yourself. The first visitor becomes the owner.

GPU, VRAM, and RAM

Open Task Manager, then Performance, then GPU. Read Dedicated GPU memory. Then open Performance, then Memory, and read the total.

Dedicated GPU memory is about 12 GB on an NVIDIA card. That is the card that was tested, and it is the recommendation. A smaller dedicated figure was not tested. A larger one was not tested either.
If the page shows shared memory and little or no dedicated memory, that is different graphics hardware than the test machine.
RAM is 32 GB, which is what the tested PC had. This page does not claim a lower amount was tried. Write your number down before you pull a model.
You are choosing a model in the 8B to 14B range. Example names from the guide are llama3.1:8b and qwen3:14b. On the tested 12 GB card, a 14B model with a 32,768-token context filled the GPU and spilled about 15 percent of the work onto the CPU. An 8B model was the easier fit.
You are not expecting an Ollama version number from this kit. Ollama was already installed on the test machine, and the log did not record its version.

Disk space

Plan on about 8 GB for the stack itself. Check free space in File Explorer, then This PC, on the drive where the stack will go. The guide's breakdown is:

About 2.0 GB for uv, including its cache.
About 2.8 GB for Open WebUI.
About 3.2 GB for n8n, including the npm cache.
Extra room for model files. They are often larger than the stack. The install log does not record a separate size for them. A download that fills the disk will stop halfway.

Put the stack in one folder: %USERPROFILE%\LocalAIStack. In PowerShell that is $env:USERPROFILE\LocalAIStack.

Tool versions and install method

These are the versions and the methods the install guide records. Newer versions may differ.

node --version prints a 24.x line. The tested version was v24.19.0. n8n 2.41.6 was already on that Node, so no second copy was installed. The package data says engines.node is >=24.0.0. The npm notes checked that day also asked for 20.19 through 24.x. Use a 24.x release if you still need Node.js, from https://nodejs.org/. The log does not record the command that printed the version.
uv 0.12.22, from the official installer, after you have read https://astral.sh/uv/install.ps1. The tested command from the stack folder is below. It does not change the user PATH.
CPython 3.11.17, installed with uv into the stack folder, then Open WebUI 0.11.4 in that virtual environment. The tested machine already had Python 3.13.2. The stack did not use it. uv python install also registers CPython 3.11.17 for your user (PEP 514). That registry key was the one extra change.
Ollama, if you do not already have it, from the Windows installer at https://ollama.com/download. No API key. Start Ollama before the chat page. You need at least one model. The tested run did not pull a model. llama3.1:8b is the comfortable example on a 12 GB GPU.
n8n 2.41.6 from a local npm install inside the stack, not a global install. The tested command is below. It added 2298 packages in about 5 minutes and exited 0.

uv, from the stack folder, in the same PowerShell window:

$env:UV_UNMANAGED_INSTALL="$env:USERPROFILE\LocalAIStack\uv\bin"; irm https://astral.sh/uv/install.ps1 | iex
$env:UV_PYTHON_INSTALL_DIR = "$env:USERPROFILE\LocalAIStack\uv\python"
$env:UV_CACHE_DIR = "$env:USERPROFILE\LocalAIStack\uv\cache"
uv python install 3.11 --no-bin
uv venv --python 3.11 open-webui\venv
uv pip install --python open-webui\venv\Scripts\python.exe open-webui==0.11.4

n8n, from the stack folder:

cd "$env:USERPROFILE\LocalAIStack\n8n\app"
npm.cmd install n8n@2.41.6 --cache ..\npm-cache --no-fund --no-audit

Open WebUI then listens on 127.0.0.1 port 8080, with OLLAMA_BASE_URL set to http://127.0.0.1:11434. The first start downloads the embedding model sentence-transformers/all-MiniLM-L6-v2. Wait until that window is quiet, then claim the admin account.

Loopback ports

Before you install, these ports should be free. After you install, the listeners should be on 127.0.0.1 only. The tested install had nothing on 0.0.0.0. Do not forward these ports to a LAN or the internet.

11434 is Ollama. The local API is http://127.0.0.1:11434. http://127.0.0.1:11434/api/tags should be a JSON model list when Ollama is running.
8080 is Open WebUI. Health check: http://127.0.0.1:8080/health returns HTTP 200 and {"status":true}. http://127.0.0.1:8080/api/version returns 0.11.4.
5678 is the n8n editor. http://127.0.0.1:5678/healthz returns HTTP 200 and {"status":"ok"}. On the tested run the editor took about 100 seconds, and the page was about 52,122 bytes after migrations.
5679 is n8n's task-runner broker. It opens after n8n is running. It is part of n8n, not a fourth program. On the tested run it was loopback only, the same as 5678.

In PowerShell, before install:

netstat -ano -p tcp | findstr ":11434 :8080 :5678"

No lines means nothing is listening yet. A line that starts with 127.0.0.1 is loopback. A line that starts with 0.0.0.0 is wider than the tested setup. Do not install a second copy on a port you do not recognize. This page does not ask you to force-stop a process by its id.

Docker and WSL

Docker is not installed, or you are leaving it unused. The tested path does not need it. The test machine did not have Docker. This kit does not include Docker commands.
The Windows Subsystem for Linux is not installed, or you are leaving it unused. The tested path does not need it. Podman was not installed either.
You are staying on n8n 2.41.6 from npm. n8n's npm install is deprecated from 3.x, and the running 2.41.6 process warned that a future version expects a container. The guide points at Docker documentation for that later move. You do not need that move to use this kit.

Before you trust a "done" claim

A sentence that says Created or Updated is a claim about a file. The failure log records 493 lines from 2026-09-26 12:57:11 through 2026-10-03 10:27:35. That span is 6 days, 21 hours, 30 minutes, and 24 seconds. 492 of those lines said a 248-byte script was created again. The file's last-write time never moved after 2026-09-26 12:56:24. The wording changed 229 times, so a check that only looks at the previous sentence would have let the reworded lines through.

Before the agent wakes, run New-Receipt.ps1 on the files it might touch. The snapshot stores whether each path exists, its size, its last-write time, and its SHA-256. A missing path is recorded as absent. The receipt is still written.
You will accept "Created" or "Updated" only through Add-CycleLog.ps1, with that snapshot. The line is kept only when every claimed file exists now and its SHA-256 is not the pre-wake hash.
You know the rejects: no snapshot (no-snapshot), unreadable snapshot (bad-snapshot), file missing (missing), same SHA-256 (unchanged or unchanged-hash, including a clock-only touch), and the same actor plus the same message as the previous line (duplicate). A rejected line is not appended to the cycle log.
ActorKind is agent, helper, or human. A required kind stops a blank label. It does not prove the writer picked the true kind.

The parameters from the scripts readme, with its placeholder paths:

powershell -NoProfile -ExecutionPolicy Bypass -File Add-CycleLog.ps1 -LogPath "%USERPROFILE%\my-agent\cycle.jsonl" -Actor my-agent -ActorKind agent -Message "Created out.txt" -SnapshotPath "%USERPROFILE%\my-agent\snapshot.json" -BaseDir "%USERPROFILE%\my-agent" -RejectLogPath "%USERPROFILE%\my-agent\rejected.jsonl"

GPU idle-unload check

gpu-guard-unload.json runs every 10 minutes and also has a manual trigger. It GETs http://127.0.0.1:11434/api/ps. Shipped config: dryRun true, idleMinutes 15, assumedKeepAliveMinutes 5, unloadPinned false.

Estimated last use is expires_at minus 5 minutes. A model is idle when now minus that estimate is at least 15 minutes. A model whose expiry is more than 24 hours away is pinned and is left loaded unless you set unloadPinned true. While dry-run is true, an idle model is only logged as would unload plus the model name. No unload request is sent.

Dry-run is still true. You have not turned it off just to see a live unload.
You have replaced the placeholder logPath with a real folder. n8n does not expand %USERPROFILE%. The shipped placeholder starts at %USERPROFILE%\LocalAIStack.
After a model job, you read the loaded-model list and the GPU memory. In the recorded incident the idle reading was 959 MiB, against a target of under 1 GB. The workflow does not change context length.
If you later set dry-run false, you know the live call is one POST to /api/generate with {"model": "<name>", "keep_alive": 0} for each idle model.

On 2026-10-03 the tested 12 GB card had qwen3:14b loaded at 12 GB, 15 percent of the work on the CPU and 85 percent on the GPU, context 32,768. Ignore a number painted on a provider name. Read the context on the loaded-model list.

The kit

This page is free. The kit is the tested path, the guardrail scripts, the workflows, and the failure log. The price is $24. One-time payment. Personal-use license. 30-day refund.

Buy the kit for $24

Free preflight checklist. No email needed.

Checkout is handled by Payhip. You get the download link from the store right after paying.

Back to the kit page