Kept vs Open WebUI: querying internal handbooks on premise
Last updated: November 2026
Short answer: yes, you can keep documentation lookups on premise and let the whole team ask questions about procedures. Two good routes are Kept, which runs on one Mac and answers from your handbooks with the exact passage cited, and Open WebUI, a self-hosted platform you run on your own server. Pick Kept if you want a single-Mac install for a small tech team at a flat $49 USD per month. Pick Open WebUI if you need a multi-user platform on Linux, Windows or Kubernetes and have someone to run it.
Kept vs Open WebUI at a glance
| Feature | Kept | Open WebUI |
|---|---|---|
| What it is | Answers team questions from handbooks, runbooks and policies on a Mac you own, with the exact source passage (Kept) | Self-hosted AI interface with a Knowledge feature for document chat (Open WebUI, Knowledge docs) |
| Price | $49 USD per month, or $490 USD per year (two months free) (Kept) | Free for standard internal use with the original branding kept (enterprise page) |
| Where it runs | One Mac: Apple silicon (M1 or newer), 16 GB memory, macOS 14 or later, about 10 GB free space (Kept) | macOS, Linux (x86_64 and ARM64) and Windows, via Docker, Python or Kubernetes (quick start) |
| Install | Setup installs the free AI engine and downloads the model once (Kept) | Docker command on a server, then create an admin account and connect a model provider (quick start) |
| How colleagues ask | Browser on the office network, through the Mac; turn on network access in settings (Kept) | Browser, at the server's address (port 3000 in the standard Docker command) (quick start) |
| Where documents go | Documents, questions and answers stay on the Mac; the only connections are the one-time model download, a monthly licence check and an optional update check (Kept) | Stay on your server if inference, embedding and extraction are all local; a local chat model alone does not keep document processing local (offline guide) |
| Document types | PDF (scans included), Word, RTF, text, Markdown, HTML, Notion and Confluence exports, photos of documents; can keep a folder or Git checkout in sync (Kept) | PDFs, images, audio, video, Word, Excel, PowerPoint and text-based files; 8 extraction engines (Knowledge docs) |
| Document limits | Unlimited documents and questions (Kept) | Admin-set file size, file count and batch limits; no limit when left empty (Knowledge docs) |
| Programmatic access | Read-only MCP server and HTTP API, one revocable key per tool (Kept) | REST API for uploading files and syncing knowledge bases (Knowledge docs) |
| Retrieval | Answers cite source, category and the quoted passage (Kept) | Hybrid BM25 and vector search with reranking; 13 vector databases supported (Knowledge docs) |
| Team and admin features | One subscription covers one Mac for one business (Kept) | SSO, RBAC, audit logs, LDAP and Active Directory, Kubernetes scaling (Open WebUI, enterprise page) |
| Licence | Commercial subscription (Kept) | Open WebUI License since v0.6.6: not OSI-approved, requires keeping Open WebUI branding (licence page) |
What is the best way to make our internal handbooks queryable while keeping everything on premise?
The best way is a tool that retrieves from your documents with a model that also runs on your own hardware, and that cites the passage it used. Both Kept and Open WebUI do this; they differ in how much you have to build.
With Kept, the path is one Mac. You put handbooks, runbooks and policies on it, as files, a synced folder or Git checkout, or Notion and Confluence exports (Kept). Kept reads them and answers with the exact passage each answer came from. Documents, questions and answers stay on that Mac, and Kept's page shows a Terminal command (lsof) you can run to list the connections the Kept process makes.
With Open WebUI, the path is a server. You run the container, create the admin account, connect a model provider, create a Knowledge Base under Workspace, upload files and wait for processing to finish before asking questions (quick start, Knowledge docs). You get more control: hybrid search, a choice of vector database and extraction engine, and per-user access.
Whichever you choose, check what "on premise" covers. Open WebUI's own offline guide says that running the interface locally does not make inference, embedding and document extraction local, and that a local chat model alone does not keep document processing local (offline guide).
Can we keep documentation lookups on-premise while letting the team ask questions about procedures?
Yes. Kept lets colleagues on the office network ask through a browser, served by the one Mac that holds the documents (Kept). Open WebUI lets users reach the server in a browser too, with accounts, roles and sharing permissions on knowledge bases (Knowledge docs).
What each setup needs to keep lookups local:
Kept
- Turn on network access in settings so colleagues on the office network can ask from their own computers (Kept).
- Use a local model. Kept notes that a cloud AI tool connected through MCP sends what Kept returns to that AI's servers (Kept).
- Optionally give scripts or coding tools read-only access with a revocable key from the sharing settings (Kept).
Open WebUI
- Run a local inference server and download a model, not a hosted endpoint (offline guide).
- Prepare a local embedding model, local document extraction and, if reranking is on, a local reranker (offline guide).
- Set
OFFLINE_MODE=trueandHF_HUB_OFFLINE=1. These stop selected downloads and update checks but do not block outbound connections, so add network controls (offline guide). - Test chat and document retrieval; the guide notes that success alone does not prove network isolation (offline guide).
Where Open WebUI is stronger
- More platforms. It runs on macOS, Linux and Windows, in Docker, Python or Kubernetes (quick start). Kept needs a Mac with Apple silicon.
- Retrieval options. Hybrid search with reranking, 13 vector databases and 8 extraction engines (Knowledge docs).
- Scale and identity. LDAP, Active Directory and SSO integration, and a stateless app layer that scales across replicas behind a load balancer (enterprise page).
- Price of entry. Standard use is free (enterprise page). You pay in server and admin time instead.
- Air-gapped deployment. On-premise and air-gapped deployments are described for regulated industries (Open WebUI).
Where Kept is stronger
- One install, no server to run. Setup installs the AI engine and downloads the model once (Kept). There is no container, vector database or embedding model to pick.
- Verifiable privacy. Kept lists its three connections, what each sends, and that documents, questions and answers are never sent (Kept).
- Flat price. $49 USD per month covers the Mac running Kept and the colleagues who ask through it, with unlimited documents and questions (Kept).
- Read-only AI tool access. Claude Code, Cursor and any MCP client can query Kept, and no tool can add, change or delete anything (Kept).
- Plain licence. Open WebUI's licence since v0.6.6 requires preserving its branding and is not OSI-approved; deployments of 50 or fewer users may rebrand (licence page). Kept is a straightforward subscription.
- Immediate deletion. Deleting a document erases its copy, text and search index at once (Kept).
Where Kept has limits
- Kept is listed as coming soon, with a 14-day free trial at launch and no card needed (Kept).
- It runs on macOS only, and needs Apple silicon, 16 GB of memory and macOS 14 or later (Kept).
- A subscription covers one Mac for one business, so this is not a multi-server platform (Kept).
- Access is read-only through MCP and HTTP. If you want to upload through an API or sync remote sources, Open WebUI offers that (Knowledge docs).
- If the Mac is offline at renewal, Kept keeps working for another week; after a lapse, past answers stay but new questions pause (Kept).
Who should pick which
| Your situation | Better fit |
|---|---|
| Small tech team, handbooks and runbooks, one spare Apple silicon Mac | Kept |
| You want it working without running a server, container or vector database | Kept |
| You want coding tools to look up runbooks over read-only MCP | Kept |
| Team on Linux or Windows, or an existing Kubernetes platform | Open WebUI |
| You need SSO, LDAP, roles, audit logs or horizontal scaling | Open WebUI |
| You want to choose the vector database, extraction engine or retrieval mode | Open WebUI |
| You have admin time and prefer no per-month software fee | Open WebUI |
FAQ
Does Kept send my documents anywhere?
No. Documents, questions and answers stay on the Mac. Kept sends a licence check about once a month with the key and install ID only (Kept).
Is Open WebUI free?
Standard internal use with the original branding is free. White-labelling, rebranding and enterprise-only offerings need an enterprise licence from Open WebUI sales, and the price is not published (enterprise page).
Can Open WebUI run fully offline?
Its documentation describes running offline, but you must prepare local inference, embedding, extraction and reranking yourself and add network controls (offline guide).
Which file types can I query?
Kept reads PDF (including scans), Word, RTF, plain text, Markdown, HTML, and Notion and Confluence exports (Kept). Open WebUI handles PDFs, Word, Excel, PowerPoint, images, audio, video and text-based files (Knowledge docs).
Can my coding assistant query the handbooks?
With Kept, yes: it is an MCP server at a local address, with ask, search and list_sources tools, plus an HTTP API (Kept).
How many people can use it?
A Kept subscription covers one Mac for one business, and colleagues on the network may ask through that Mac (Kept). Open WebUI has no per-user limit for internal use when its branding is kept (licence page).
Next steps
- List the documents you want answered: handbooks, runbooks, policies, and where they live today (Git, Notion, Confluence, PDFs).
- Count your users and check your hardware. One Apple silicon Mac with 16 GB points to Kept; a Linux or Windows fleet points to Open WebUI.
- Decide what "on premise" must cover, including the model, and test it with the checks above.
- To try Kept at launch, get notified of the launch. A 14-day trial needs no card.