The private and local stack

All three tools here have a genuine free tier, so a working configuration costs nothing while you evaluate it. The 03 COSTS table above breaks down each component. The cost drivers here are hardware and ownership: the machine that runs the models, the server that hosts the team front end, and the time to keep both patched and updated. Model quality is bounded by the machine underneath, so budget capable kit and the attention to maintain it as much as any spend.

No. Ollama plus Open WebUI is a complete private chat capability (Ollama's local runtime needs no account or cloud; it also offers an optional paid cloud tier if you want hosted models), and Open WebUI carries document Q&A too, so many teams stop there. The numbered steps double as the adoption order: stand up Ollama first, since two commands prove the hardware; add Open WebUI when the runtime needs accounts and admin controls to be shareable; bring in AnythingLLM last, pointed at the same backend, once citation-backed retrieval becomes a daily need.

Then you already have the engine: Ollama serving open models locally behind a standard API, the hard part proven. What it does not give you is a team. This stack keeps Ollama and builds on top: Open WebUI, source-available and free to self-host, wraps it in accounts, roles and an admin panel, turning one person's runtime into shared infrastructure; AnythingLLM adds workspace isolation and citation-backed retrieval for serious document work. The prototype becomes the platform.

Open WebUI and AnythingLLM both answer questions from your documents, the one real overlap here. The dividing rule is how central the documents are. Open WebUI wins when files are occasional context alongside general chat. AnythingLLM wins when documents are the workload itself, where workspace isolation keeps clients and projects apart and every answer cites its source. The AnythingLLM and Open WebUI comparison, linked above, covers the finer calls.

Two signals, both about scale, not spend. The first is concurrency: when enough people hit the stack at once that a single-machine runtime stalls, you outgrow one box serving one queue. The second is production duty: when the capability must stay up with the reliability of a real service, monitored and highly available rather than a server someone reboots. Both point to the world-class tier, the production private AI stack.

Almost anything, which is the whole point of building it this way. Unusually for this taxonomy, no paid tier gates the sensitive work: the floor is configuration, not licensing. Kept self-hosted against the local backend, the stack leaks nothing, so confidential documents, client files and internal data are fair game. The one discipline is guarding the exits: no cloud model key in Open WebUI or AnythingLLM, Ollama's web features off, AnythingLLM self-hosted not hosted. Open source makes the guarantee auditable.