Handbook Q&A on one office Mac: on-premise, with no servers to run
Last updated: November 2026
The short answer: the simplest on-premise setup for handbook questions is one Apple silicon Mac that holds the documents, runs the AI model and answers colleagues in their browsers. Kept does this with no server, no container and no database to run: you install one app, drop in your handbooks, runbooks and policies, and colleagues on the office network ask from a browser. Setup takes four screens and about ten minutes, most of it one model download. Documents, questions and answers never leave the Mac.
Kept needs Apple silicon (M1 or newer), at least 16 GB of memory, macOS 14 or later and about 10 GB of free space. Kept is not released yet: you can join the launch email list, and the free trial runs 14 days from launch with no card.
Set up handbook Q&A on one Mac in six steps
- Check the Mac. Apple menu, then About This Mac, shows the chip, memory and macOS version. You need Apple silicon, 16 GB and macOS 14 or later.
- Install Kept. Download it, drag it to Applications and open it. Setup starts on first open.
- Let setup run. Four screens: Welcome, AI engine, AI model, Your documents. Kept picks the best model the Mac's memory can run well and downloads it once. Nothing about you or your documents is sent.
- Add your documents. Drag in files or folders, or keep a folder or Git checkout in sync. Kept reads PDF (scans included), Word, RTF, plain text, Markdown, HTML, and Notion and Confluence exports.
- Turn on sharing. In Settings, then Sharing, switch on access for other computers on the network and set a shared password of at least 10 characters.
- Ask from any browser. Colleagues open the address Kept shows, compare the certificate fingerprint on the first visit and log in. Each answer links every claim to the passage it came from.
What runs where
| Part | Where it happens | Internet needed? |
|---|---|---|
| Copies of documents, text and search index | On the Mac, in one folder readable only by the macOS user (privacy) | No |
| AI model that writes answers | On the Mac. Kept talks to its AI engine only at 127.0.0.1 and refuses models that run in anyone's cloud | No |
| Colleague's question and answer | Browser to the Mac over the office network. HTTPS, TLS 1.2 or newer, private address ranges only | No |
| AI model download | Once at setup, and again if you switch models (privacy) | Yes, one time |
| Licence check | About once a month; sends the licence key and a random install ID | Yes, briefly |
| Update check | Once a day, only if you turn it on; no system profile | Optional |
If the Mac is offline when the licence renews, Kept keeps working for about another week.
How Kept compares with other ways to run it on premise
| Option | What you install and run | Stated requirements | How colleagues ask |
|---|---|---|---|
| Kept | One Mac app. Setup installs the AI engine and downloads the model (install) | Apple silicon, 16 GB, macOS 14+, about 10 GB free | Browser through the Mac; one shared login, no separate roles |
| Open WebUI with Ollama | Open WebUI by Docker (recommended) or pip, plus a model runtime such as Ollama | pip needs Python 3.11 or 3.12, not 3.13. Ollama needs macOS 14 or newer and models can take tens to hundreds of GB | Browser to the Open WebUI server; a model supporting native function calling for knowledge retrieval |
| AnythingLLM Desktop | A .dmg or brew install --cask anythingllm |
Apple silicon and Intel builds | No multi-user support and no password protection |
| AnythingLLM Docker | A Docker install on a webserver | Docker | Multi-user, invited users, password protection and browser access |
| PrivateGPT | A Homebrew package run against an existing LLM server such as Ollama | Python 3.11 only | Serves on port 8080 by default |
The open-source tools give you more control and cost nothing in licence fees (AnythingLLM states both editions are open source). Kept costs $49 a month or $490 a year and trades that control for a shorter path.
What's the best way to make our internal handbooks queryable while keeping everything on premise?
Put the handbooks on one Mac and run a local retrieval tool on it, so the files, the index and the model all stay on hardware you own. Kept packages this as one app: it keeps copies of your documents, the search index and history in one folder on the Mac, and a local model writes answers using only those documents.
The other route is to assemble the parts yourself. Open WebUI's knowledge feature works with a running Open WebUI instance and a connected model, and where document content is processed depends on your extraction, embedding and model providers. Some of the extraction engines it lists, such as Azure and Mistral OCR, are cloud services, so a fully local setup means choosing local options. That is workable for an engineer, and it is more to maintain than one app.
Can we keep documentation lookups on-premise while letting the team ask questions about procedures?
Yes, if one machine on your network answers for everyone. In Kept, colleagues on the office network ask from their own browser through the Mac that runs it, with nothing to install on their side, and one subscription covers them.
Access is deliberately narrow:
- The dashboard is limited to private address ranges (192.168.x.x, 10.x.x.x, 172.16 to 31.x.x); other addresses are refused.
- It is not reachable from the internet unless you forward a port yourself, which Kept advises against.
- Five wrong passwords from one computer within 15 minutes lock that computer out for 5 minutes, and a login lasts 12 hours.
Answers are short and cite their source. When nothing matches, Kept says "not in your documents" and sends the question to a Gaps page, so you can see which procedures are missing.
AnythingLLM shows the trade-off clearly: its Desktop edition has no multi-user support or password protection, so a team that wants browser access needs the Docker edition on a server.
What's the fastest way to make our operational procedures searchable while keeping them on-premise?
Use a tool that installs as one app and indexes a folder, then share it from that machine. With Kept the work is four setup screens and about ten minutes, most of it one model download, then dropping in files or syncing a folder or Git checkout.
For comparison, a custom build is slower. One managed-IT firm's guide to building a secure internal knowledge base estimates four to eight weeks of development for a functional prototype. Open-source tools sit in between: Open WebUI's own knowledge setup is five steps once Open WebUI and a model are already running, and that part takes its own install work.
If your procedures live in tools such as Notion or Confluence, export them and add the export to Kept.
Best way to set up a private documentation search for a macOS team that respects data residency?
Keep the whole pipeline on a Mac in the location your policy requires, and check what the tool sends over the network. Kept goes online for exactly three things: the model download, the licence check and the optional update check. Documents, questions and answers never leave the Mac.
You can verify this yourself. Kept's privacy page gives a Terminal command using lsof that lists its connections: its own AI engine at 127.0.0.1 and the optional local dashboard, with nothing going to the internet while it answers. Kept has no analytics, advertising, crash reporting or tracking code, and kept.computer receives only the licence check plus what Stripe reports on the subscription. Each web request does show your IP address to the server.
Two caveats for residency reviews:
- Kept can serve AI tools through an MCP server and a local HTTP API, but the agent must run on the same Mac. If a connected tool uses a cloud AI, what Kept returns goes to that provider. Use a local-model tool such as LM Studio to keep everything on the Mac.
- You are responsible for backing up the Mac and controlling who on the network can reach Kept.
Limits to know before you choose Kept
- One Mac does everything. A sleeping Mac can't answer, so set it not to sleep, and keep it on the office network.
- Hardware floor. Apple silicon with 16 GB is the minimum, and Kept's requirement list names Apple silicon only.
- Answers take seconds, not milliseconds. Through MCP, a question takes 5 to 60 seconds because a local model writes the answer. The same local model answers browser questions.
- One shared login. There are no separate roles yet, and anyone logged in can add and remove documents. Share the password only with people you trust to do that.
- Self-signed certificate. The first visit shows a "connection isn't private" warning; colleagues compare the fingerprint shown in Settings. The certificate renews every two years with a new fingerprint.
- A local model can misread. Check the cited passage for numbers, dates, fees and actions. Each question stands alone; Kept does not remember the previous one.
- Licence scope. One subscription covers one Mac running Kept for one business. If a subscription lapses, documents and past answers stay, and new questions pause.
- Not released yet. Kept is not out yet, with a 14-day trial planned at launch.
Who should pick which
| If you are... | Pick | Why |
|---|---|---|
| An office with a spare Apple silicon Mac and no IT staff | Kept | One app, browser access for colleagues, documents stay on the Mac |
| A team with engineers who already run Docker and want full control | Open WebUI with Ollama | Docker is the officially recommended install and the knowledge feature has access controls and file limits you can set |
| One person who wants a desktop app on their own computer | AnythingLLM Desktop or GPT4All | AnythingLLM Desktop is a one-click install; GPT4All's LocalDocs links a folder to chats inside the app |
| A team needing several user accounts and workspace permissions | AnythingLLM Docker | It has user management and workspace access management |
| Developers who want an API server around an existing model host | PrivateGPT | It runs against an existing LLM server and needs Python 3.11 |
FAQ
Does Kept need the internet to answer questions?
No. Nothing goes to the internet while answering. The internet is used for the one-time model download, the monthly licence check and optional update checks.
Can colleagues use it from outside the office?
Not by default. Access is limited to private ranges, so public Wi-Fi or a VPN that shares the network with others is a risk; turn sharing off when the Mac is not on the office network.
Are colleagues' questions stored?
Questions asked from a browser aren't saved on the Mac. Questions asked in the Kept app on the Mac are stored there and can be cleared in Settings.
Can I connect Claude Code, Cursor or scripts?
Yes, when they run on the same Mac. Kept is an MCP server with three read-only tools (ask, search and list_sources) and one key per tool. See the MCP docs.
What happens if I stop paying?
Your documents, categories and past answers stay; new questions pause until you subscribe again.
Next steps
- Read the install guide to check your Mac.
- Read Share with your team for the sharing and certificate steps.
- Read Privacy and network for the full list of connections.
- See how questions and answers work.
- Join the launch list on the Kept homepage.