AnythingLLM alternatives: what actually migrates when you switch

The right AnythingLLM alternative depends on which of its three editions you are leaving — Desktop, self-hosted Docker, or Cloud — because they fail for different reasons. Open WebUI and LibreChat are the closest like-for-like swaps for multi-user chat over documents; RAGFlow is the one to pick when retrieval quality is the complaint; Onyx fits connector-driven company search; LM Studio and Jan fit a reader who only ever wanted a local chat window.

The part the brand lists skip is the bill for moving. AnythingLLM exports chat history and nothing else, no alternative here documents an importer for it, and the tool you are leaving is not in trouble — Mintplex-Labs/anything-llm is MIT, not archived, with commits on its default branch as recently as September 17, 2026 and v1.16.1 published August 27, 2026 (GitHub API, re-checked September 21, 2026). So you have time to do this properly.

Why people leave, and which edition they are leaving

Four complaints account for most of it, and each one is a documented property of a specific edition rather than a general flaw.

Read from the cloud limitations page, the desktop overview and the pricing page on September 19, 2026. Two naming traps before you compare anything. AnythingLLM's own surfaces disagree on tier names — the pricing page says Basic and Pro, the docs say Starter and Professional, and nothing reconciles them — so confirm which one your account is on. And "Pro" names two unrelated products: Cloud Pro at $99/monthly, and AnythingLLM Desktop Pro, a licence key for Desktop v1.15.0 and later. The Desktop Pro page describes its features but publishes no price, and anythingllm.com/pro returns 404, so any figure you see attached to that name came from somewhere else. No seat count, storage quota or usage quota is published for Cloud Basic or Pro either.

Note what is not on that list: cost of the models themselves. That bill does not go away whichever self-hosted tool you switch to, because each of them expects you to supply the key. It is not identical, though — retrieval defaults differ between these products, and they decide how much context every question carries.

Which alternative fits which job

Prices and repository metadata checked September 19, 2026 against Open WebUI's enterprise page, librechat.ai, onyx.app/pricing, ragflow.io, dify.ai/pricing, LM Studio's licence post, msty.ai/pricing and the GitHub API. Three cautions. RAGFlow's $29 and $129 are promotional figures displayed struck through from $59 and $259, with no stated end date, and the page never defines what one of its "credits" buys. Open WebUI's branding clause, spelled out in its licence page and the repository LICENSE, is the detail most people miss: keeping the branding intact is the free path, and removing or altering it is permitted only where the deployment has fifty or fewer end users in any rolling thirty-day period, where you hold specific prior written permission from the copyright holder, or under an enterprise licence. And Onyx is a badly collided name — only onyx.app can supply Onyx prices; a search without the domain returns a macOS maintenance utility, a printing RIP vendor and several other companies.

Four projects that are not what older articles say

Alternatives lists rot faster than the products on them. These four are still recommended widely and should not be shortlisted without knowing their current state, all checked September 19, 2026.

One more to treat carefully: Quivr's repository now resolves to a different GitHub organisation, and its README describes a RAG core library rather than an end-user application. No dated announcement explaining the move was found, so that is a fact about the redirect, not about an acquisition.

What migrates, and what does not

This is the cost that should decide your shortlist, and it is the same for every destination on this page.

What comes with you: chat history, in exactly four formats — CSV, JSON, JSONL and JSON (Alpaca) — served from AnythingLLM's /system/export-chats endpoint. That is the complete list of export types in the shipped source.

What does not: everything else. AnythingLLM's documentation was searched for a document export, a vector export, or a documented backup and restore procedure, and none was found. That is absence of evidence rather than proof of impossibility — but you cannot plan a migration around an undocumented path. Workspaces, embeddings, and the documents themselves have no documented route out. The Community Hub's import strings move agent skills, system prompts and slash commands — the three item types its docs list as supported — but only between AnythingLLM instances, and on self-hosted installs skill imports fail unless COMMUNITY_HUB_BUNDLE_DOWNLOADS_ENABLED is set. No alternative on this page documents an importer for AnythingLLM data; that is what I checked for, in each product's own docs. LibreChat is the only one that documents a conversation importer at all, and the sources it names are ChatGPT, Claude and ChatbotUI v1 — not AnythingLLM.

So migration means re-ingestion, and re-ingestion has one hard prerequisite: your original source files must exist outside the app. Practically:

1. Export chat history first, while the instance is still running. It is the only export AnythingLLM has, and it is the record of what people actually asked — useful later for judging whether the replacement answers as well.
2. Gather the original documents from wherever they came from — the shared drive, the wiki, the ticket system — not from inside AnythingLLM.
3. Copy the Desktop storage folder to cold storage before uninstalling anything. The documented locations are /Users/<usr>/Library/Application Support/anythingllm-desktop/storage on macOS, ~/.config/anythingllm-desktop/storage/ on Linux and C:\Users\<usr>\AppData\Roaming\anythingllm-desktop\storage on Windows, and anythingllm.db inside it is the SQLite database. Treat this strictly as a backup: the docs give no migration, backup or restore instructions, and whether that folder can be moved to another machine or into Docker was not tested here.
4. Stand the replacement up alongside the old one and re-ingest. Then ask both instances ten questions whose answers you already know, and compare. Retrieval defaults differ between these products, so the same corpus does not guarantee the same answers.
5. Roll back by doing nothing. Keep AnythingLLM running until step 4 passes; AnythingLLM stores vectors in an on-instance LanceDB by default, so the old index is sitting on your own disk, not in a service you have cancelled.

Pointing the new client at a model endpoint

Each of the seven products below takes a third-party endpoint in a provider slot of its own, and each has a different gate. The slot is OpenAI-compatible in every case; Jan and LibreChat will also speak the Anthropic Messages format if you ask them to. LM Studio and Msty Studio are absent from this table on purpose — this page did not establish a documented remote-endpoint path for either. This is the reconfiguration step, and knowing the gate in advance is what keeps it short.

Each gate above is quoted from the product's own setup page: AnythingLLM's OpenAI (generic) provider, Open WebUI's OpenAI-compatible guide, LibreChat's custom endpoint reference, Onyx's custom inference provider page, RAGFlow's model API key setup, Jan's custom endpoints page and PrivateGPT's README.

One boundary worth settling before you configure anything. In AnythingLLM, LibreChat, RAGFlow, Dify and Cherry Studio the embedder is registered as a separate provider from the chat model, with its own base path and its own key. Kunavo serves no embedding model and no speech-to-text model, so pointing an embedder at it returns a "model is not available" error. Chat completions are the only stage of a RAG pipeline Kunavo can answer. Leave the embedder on the bundled local model, or on a separate vendor:

# Chat goes to an OpenAI-compatible endpoint.
LLM_PROVIDER='generic-openai'
GENERIC_OPEN_AI_BASE_PATH='https://api.kunavo.com/v1'
GENERIC_OPEN_AI_API_KEY='sk-...'
GENERIC_OPEN_AI_MODEL_PREF='claude-haiku-4-5'
GENERIC_OPEN_AI_MODEL_TOKEN_LIMIT=200000

# Embeddings do NOT. Kunavo serves no embedding model, so the
# embedder stays on the bundled local one (or another vendor).
EMBEDDING_ENGINE='native'
EMBEDDING_MODEL_PREF='Xenova/all-MiniLM-L6-v2'

That split is not an AnythingLLM quirk: LibreChat's RAG API has its own EMBEDDINGS_PROVIDER and its own RAG_OPENAI_BASEURL, and RAGFlow, Dify and Cherry Studio each register embedding models in a different place from chat models. Budget the two halves separately too. For the retrieval side of the design, see RAG implementation; for what the compatible surface does and does not cover, OpenAI-compatible API.

What a month of chat costs after you switch

The client is free in almost every row above, so the recurring number is tokens. This is illustrative token arithmetic, not a measured cost and not a bill ceiling. Assume a small team asking 1,200 questions a month, where each turn sends 6,000 input tokens (retrieved chunks, system prompt and a little history) and returns 500 output tokens — 7.2M input and 0.6M output per month. Rates are live Kunavo catalog prices per million tokens.

Two readings. First, at this shape the model choice moves the bill by a factor of about 5.0 between the cheapest and most expensive rows, while the retrieval settings that decide how many chunks land in those 6,000 input tokens are set in the client, not the provider — so tune the client before you shop for rates. Second, put it beside the hosting figures: a $50/monthly Cloud plan buys hosting and still requires you to bring the key, so the comparison is hosting plus tokens against your own Docker host plus the same tokens.

Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. Cache charges, image or video calls and anything your embedder costs elsewhere sit outside this example. The minimum top-up is $10 in prepaid credit — a funding minimum, not a task fee or a subscription. See billing details.

Setting up the replacement

If Open WebUI is where you land, Kunavo publishes a configuration reference for it: the Open WebUI setup guide. That is a published setup document, not a compatibility test — none of the clients on this page has been runtime-tested against Kunavo's endpoint, and everything above was read from each project's own documentation, licence files, published source and repository metadata. Keep a working route available while you try it, run one bounded task, then read the charge your account actually recorded for it. Create a Kunavo account when you are ready to fund a key.

Still deciding between two specific tools rather than the whole field? AnythingLLM vs Open WebUI takes the closest pair head to head, and LibreChat vs Open WebUI covers the other common shortlist. AI cost optimization covers trimming the token side once you have chosen.

FAQ

Vendor pricing pages, product documentation, licence files, published source and repository metadata for every project named above checked September 19, 2026 and re-checked September 21, 2026, first-party only. Kunavo token rates come from the live catalog, and every dollar example here is illustrative token arithmetic rather than a measured cost. No client on this page has been runtime-tested against Kunavo's endpoint.