Honest-Agent Starter Kit
Free Local AI Agent on Windows
Setup + guardrails + lessons. Not a magic autonomous employee, and this page says so.
"Free" in the line under the product name means the model stack runs on your own Windows PC, without a model-API subscription. The kit that teaches the tested path is a paid download.
You get a from-scratch Windows setup for someone who is not a developer: Ollama, Open WebUI, and n8n, on a path that was tested without Docker and without the Windows Subsystem for Linux. You get PowerShell scripts that check whether a "done" claim is true. You get an n8n workflow you can import that can free a GPU from an idle model. You get a real failure log from a local setup.
One-time payment. Personal-use license. 30-day refund.
Free preflight checklist. No email needed.
Tested versions and hardware are below. The preflight checklist is free. No email needed.
This kit is the install path, the checks, and the write-up of what broke. It is not a fully autonomous agent. Nothing in the download runs your accounts, sends messages on its own, or stays up after you close it. You start the stack when you want it.
The install guide starts from an empty stack folder. Ollama runs the model. Open WebUI is the chat page in your browser. n8n runs workflows. The tested path does not ask you to install Docker or WSL. When n8n moves to 3.x, its npm route is deprecated, and the guide points to Docker for that later step. This download stays on the path that was tested: n8n 2.41.6 from npm.
The scripts snapshot files, then check a later claim of "Created" or "Updated" against that snapshot. A claim passes when the file is new or actually changed. A repeat of the same claim, or a claim with nothing to compare, is rejected. That catches a false "done" about a file. It does not promise that a model never says something false.
One workflow asks Ollama which models are loaded, estimates which have been idle, and can unload those. It ships in dry-run: it records what it would unload, and it does not unload anything until you turn dry-run off. You import the file into your own n8n. You decide when a live unload is acceptable.
The failure chapter is a record from a local setup: claims of done that were not true, a model that sat on the GPU, and the other breaks that followed. Each section says what in this kit addresses it, or says plainly when the kit does not. There are no ratings and no quotes from buyers on this page.
These are the files in the download. Read the install guide first, then the rules, the scripts, the workflows, and the failure log.
01-install-guide/INSTALL-GUIDE.md
Windows steps for Ollama, Open WebUI, and n8n on the tested path, including start scripts and health checks.
02-rules/AGENTS.template.md
House rules you copy to AGENTS.md, with blanks for the agent name, the owner, and where you approve sends.
03-scripts/New-Receipt.ps1
Writes a JSON snapshot of files: exists or not, size, last write time, and SHA-256.
03-scripts/Test-Claim.ps1
Checks one file against an earlier snapshot and passes only if that file is new or changed.
03-scripts/Add-CycleLog.ps1
Appends a log line for a Created or Updated claim only when the file really changed, and rejects duplicates.
03-scripts/Run-SelfTests.ps1
Runs the script checks in a temporary folder and exits 0 only when every case passes.
03-scripts/README.md
Explains the snapshot, wake, claim, and verify flow, with exit codes and example commands.
04-n8n-workflows/daily-draft-approval.json
An importable daily draft workflow that checks the text and waits for your approval, with the send step shipped turned off.
04-n8n-workflows/file-change-verifier.json
An importable workflow that hashes a folder and writes a receipt of added, changed, unchanged, and removed files.
04-n8n-workflows/gpu-guard-unload.json
An importable workflow that estimates which local models are idle and can unload them, with dry-run left on.
04-n8n-workflows/README.md
What each workflow does, how to import it, and what you edit before you rely on it.
05-what-failed/WHAT-FAILED.md
A real failure log from a local setup, including false done claims and a model that held the GPU, and what in the kit addresses each one.
README.md
Start here: what the kit is, the order to read it, and a map of the files.
THIRD-PARTY.md
The free projects you download yourself, with each project's own license. None of their code is in the kit.
LICENSE
Personal use for one buyer. You may change the files for yourself. You may not redistribute them.
Tested on Windows 11, 2026-10-03: Open WebUI 0.11.4, n8n 2.41.6, Node.js 24.19.0, uv 0.12.22
Ollama was already on the test machine. Its version was not recorded, so this page does not name one. On that 12 GB GPU, a 14B model with a very large context filled the card and spilled about 15 percent of the work onto the CPU. An 8B model was the easier fit. A smaller GPU was not tested.
Ollama, Open WebUI, n8n, uv, and Node.js are free and open projects. You download them from their official sites. The kit does not include their programs. The kit sells the tested path, the guardrail scripts and workflows, and the lessons.
The price is $24. You pay once. There is no subscription in this offer.
The license is personal use for a single buyer. A 30-day refund comes with the purchase.
Paying once is not a promise of updates on any schedule. This page sets no update period.
Free preflight checklist. No email needed.
Before you install, check the GPU, the RAM, the free disk, Windows, Node.js, and the local ports. The checklist is one markdown page next to this file. No email needed. There is no signup form on this site.
No. This kit is setup, guardrails, and lessons. A local model that chats on your PC is not an employee that runs itself. You start the programs. You approve anything that would send a message. The scripts check file claims. They do not give the model permission to act on its own, and they do not promise that it never says something false.
The tested path does not need Docker or the Windows Subsystem for Linux. That path is Ollama, Open WebUI, and an npm install of n8n 2.41.6. n8n's own notes say the npm route is deprecated from n8n 3.x, and the running 2.41.6 process warned that a future version expects a container. The install guide points to n8n's Docker documentation for that later move. This kit does not include Docker steps.
No. The tested path, the start scripts, and the PowerShell checkers are for Windows 10 and Windows 11 only.
Only a 12 GB NVIDIA GPU was tested, with 32 GB of RAM, running models from 8B to 14B. This page does not claim a smaller card works, and it does not claim a larger card was tried. On the tested 12 GB card, a 14B model with a very large context filled the GPU and spilled about 15 percent of the work onto the CPU. An 8B model was the easier fit. Read the dedicated GPU memory figure in Task Manager before you buy. The preflight checklist shows where that number is.
Ollama, Open WebUI, n8n, uv, and Node.js are free projects. You download them from the official sites in What's free. The $24 is for this kit: the tested Windows path, the guardrail scripts, the importable workflows, and the failure log.
You have 30 days after purchase to ask for a refund. The payment is one-time.
Checkout is not open on this page yet, so the steps for asking are not listed here.
Ask within 30 days by replying to the receipt email from .
Personal use for a single buyer. You may change the files for your own use. You may not redistribute, share, resell, sublicense, or publish the kit or substantial parts of it. The free programs you install stay under their own licenses, which THIRD-PARTY.md names.
On the tested path, Ollama, Open WebUI, and n8n bind to 127.0.0.1. Other computers are not meant to open them. You choose any connection to the outside after that, such as a send step you turn on yourself. The draft workflow ships with that send step turned off. The install does download files once: the programs, an embedding model Open WebUI fetches on first start, and any chat model you pull. Those requests go to the sources you use for the download.
Checkout is not open yet.
This page has no payment form and no payment link. The buy button stays disabled until a checkout address is added.
Checkout is handled by . You get the download link from the store right after paying.