LibreChat alternatives — choosing by what you have to operate

LibreChat alternatives are an operations decision, not a price decision: LibreChat is MIT-licensed and sells nothing, so the reasons to leave are the six services its official compose file runs, an AGPL-licensed admin panel inside an otherwise-MIT project, and the fact that it has no bulk export. Match the replacement to what you are willing to operate — Open WebUI for one container, Jan or Cherry Studio for a desktop app, LobeHub Cloud or AnythingLLM Cloud for hosted, TypingMind or Msty for a bought licence — and check the migration path before you move, because the export is the part that traps people.

Two things this page will not do. It will not tell you LibreChat is dying: ClickHouse announced the acquisition on November 4, 2025 and the GitHub API still shows danny-avila/LibreChat unarchived, MIT, 44,528 stars, last pushed September 21, 2026 (re-read September 21, 2026). And it will not rank the field on features, because nothing here was benchmarked.

Why people leave a client that costs nothing

The footprint. LibreChat's official docker-compose.yml defines six services: api, admin-panel, mongodb (mongo:8.0.20), meilisearch (v1.35.1), vectordb (pgvector 0.8.0 on Postgres 15) and rag_api. Those extra containers are what give it full-text conversation search and document retrieval, so this is a trade rather than bloat — but three of those services are datastores with their own volumes to back up and upgrade (MongoDB, Meilisearch and the pgvector Postgres), on top of the api container's own. Open WebUI's own schema reference describes a single SQLite file at data/webui.db carrying users, chat history and file storage. That is the real difference between the two deployments, and it is verifiable without running either.

The licence is not uniform. LibreChat is described everywhere as MIT, and the main repository is. The Admin Panel that ships as a service in that same compose file lives in a separate repository, LibreChat-AI/admin-panel, and the GitHub API reports its licence as AGPL-3.0 (146 stars, pushed September 15, 2026). The Admin Panel documentation carries no licence statement at all; it does say the panel runs as a separate service and that you can opt out by removing the admin-panel service or gating it behind a Compose profile. If your organisation reviews licences before deployment, that is the line to bring to them — and opting out is a change to the compose file, not a fork.

The acquisition, stated plainly. ClickHouse's follow-up post commits in writing that "LibreChat remains 100% open-source under its existing MIT license" and that "for existing LibreChat deployments: nothing changes". The same post is where the only commercial change appears: the paid LibreChat Code Interpreter API was announced for discontinuation "in its current form", with the timeline set "targeting May 1st, 2026". That date has passed, and the current Code Interpreter documentation describes an Apache-2.0 service you self-host, with no price. Whether the hosted API has actually stopped billing existing subscribers was not verified here — confirming it would have meant signing in to an account, which this research did not do.

Release cadence, with the disagreement shown. The changelog dates the current stable, v0.8.7, to June 23, 2026, while the GitHub releases API shows it published June 24, 2026 and flags it prerelease: true even though the changelog calls it stable — that flag is set on every release in the list, RC or not, so it does not distinguish them. Trust the changelog for stable-versus-RC status. The head of the release train is v0.8.8-rc3, published September 15, 2026.

Which alternative fits which reader

Sorted by what you have to operate. Licences and repository figures read from the GitHub API and each project's LICENSE file on September 19, 2026; stars are a measure of attention, not of installs or team adoption.

One naming trap worth writing down: LobeChat was renamed LobeHub in a discussion opened by its founder on November 3, 2025, so pre-2.0 LobeChat tutorials and their environment-variable assumptions are obsolete. Kunavo's own LobeChat setup page still carries the old name. Cherry Studio's provider-settings documentation could not be loaded during this research, so its row in the next table is marked unverified rather than described; its licence terms below come from the repository README instead.

How each one takes your own key

This is where the alternatives actually differ, and it is the part a feature table misses. Each client in the table below documents a path for your own key and base URL, so the question is what that path will accept and what it gives up. Chatbox is the exception worth flagging early: its published plans sell bundled model access rather than the app, and its pricing page does not state whether the free tier accepts a custom endpoint.

Sources for that table, read September 19 and re-checked September 21, 2026: LibreChat's custom endpoint schema, Open WebUI's OpenAI-compatible provider guide and its Anthropic page, Jan's custom endpoint page, LobeHub's provider docs, AnythingLLM's Generic OpenAI page, which opens with its own warning that the provider is developer-focused and highly configurable, ChatWise's custom-provider page, and TypingMind's worked custom-model walkthrough.

Three boundaries worth carrying into the decision. First, LibreChat appends the completions path to your baseURL unless you set directEndpoint: true, so a base URL that is already the full completions endpoint needs that flag. Second, a custom base URL in LibreChat is not switched to the OpenAI Responses API automatically: LibreChat's OpenAI page says outright that Azure OpenAI, OpenRouter and custom or reverse-proxy base URLs are not switched automatically, so plan on Chat Completions unless you have confirmed otherwise in your own deployment. Third, agent tools — code interpreter, file search, file context, web search, MCP, OpenAPI actions and utilities, per the agents documentation — are the capability list to check against whichever client you move to, because the names and scopes are LibreChat's own. That documentation does not state what happens on a model that cannot call tools, so do not assume either a fallback or a hard failure.

One capability with no counterpart in any documentation read for this page, and the strongest argument for staying: LibreChat can price and meter your custom endpoint per model. Per the token-usage documentation, tokenConfig takes prompt, completion and context per model, with optional cacheRead and cacheWrite for providers that report cached input usage, expressed as per-million-token rates; for agents that use multiple endpoints, LibreChat uses the matching endpoint and model token config when recording usage and cost. Pair it with the balance object on the same page — enabled, startBalance, autoRefillEnabled, refillIntervalValue, refillIntervalUnit, refillAmount, which replaced the older CHECK_BALANCE and START_BALANCE environment variables — and you have per-user credit control inside the client. Rebuilding that elsewhere is real work.

version: 1.3.16
endpoints:
  custom:
    # The escape hatch a plain OpenAI-compatible client does not have:
    # provider: anthropic sends the NATIVE Messages request to your own
    # baseURL. Point baseURL at the API root, not at /v1/messages.
    - name: "Kunavo (Anthropic protocol)"
      apiKey: "${KUNAVO_API_KEY}"
      baseURL: "https://api.kunavo.com"
      provider: "anthropic"
      models:
        # models.fetch is NOT used for a native Anthropic custom endpoint,
        # so the picker is exactly what you list here.
        default: ["claude-sonnet-4-6", "claude-opus-5", "claude-haiku-4-5"]
      titleConvo: true
      titleModel: "claude-haiku-4-5"
      # Without tokenConfig, LibreChat's cost breakdowns, balance
      # transactions and agent billing have no rates for these models.
      tokenConfig:
        claude-sonnet-4-6:
          prompt: 1.2
          completion: 6
          context: 1000000
        claude-haiku-4-5:
          prompt: 0.4
          completion: 2
          context: 200000

That configuration is a published reference shape, not a compatibility test. Kunavo has not runtime-tested LibreChat or any client on this page against its endpoint; every behaviour above comes from the project's own documentation. Validate it in a staging deployment before you rely on it.

What migrates, and what you rebuild by hand

The export asymmetry is the whole migration story. LibreChat's own documentation states that "there is no built-in bulk export of every conversation at once". Export is per conversation, from the Export and Share menu, in md, txt, json, csv or screenshot; only json and csv carry all message branches, and a brand-new conversation or a search-results view has nothing to export, so the menu does not appear for it. Import, under Settings → Data Controls, accepts exactly three sources: ChatGPT, Claude and ChatbotUI v1.

Open WebUI goes the other way: Settings → Data Controls → Export Chats downloads every conversation as one JSON file, and Import Chats accepts its own native JSON, ChatGPT archives it auto-detects and converts, and custom JSON matching its documented structure. So you can leave Open WebUI in one click and leave LibreChat one conversation at a time. No official converter between the two formats was found, and an unofficial third-party importer that surfaced in search was not verified against LibreChat output — do not plan a migration around it.

Backup and rollback, in order. Copy what the official compose file actually persists before you change anything: MongoDB's store is the ./data-node bind mount, Meilisearch's is ./meili_data_v1.35.1, the vector database uses the named pgdata2 volume, and the api container has the named librechat-data volume plus ./uploads, ./images and ./logs. Take librechat.yaml and .env with them. Bring the replacement up on a different port and leave the LibreChat stack running — both are containers, and running them side by side for a week costs only host resources. Export the conversations that matter, import what the new client will take, and use the new one for real work before you decommission anything. Delete the old stores only once you have re-read a sample of the imported conversations and confirmed the content survived. If the new client cannot take a format you need, the old stack is still up, which is the entire point of doing it in this order.

What each one costs, and what you actually pay

Published prices, read September 19 and re-checked September 21, 2026. Two of these carry discount banners; where a page showed a struck-through figure, both numbers are printed. None of the self-hosted figures include your host, and no hosting bill was verified for anyone on this page — price your own host from the container count above.

Anything a reseller charges is not the project's price. One-click deploys on marketplace and PaaS listings carry their own hosting fees, and a marketplace VM image titled "LibreChat" is published by a third party rather than by the project — that listing could not be read during this research, and its fee would be the vendor's in any case.

The model bill follows your key, not your client

When you run a client on your own key, switching clients does not change token costs — the exceptions above are the tiers that sell bundled model access instead, LobeHub Cloud's credits and Chatbox's compute points, where the client and the model bill are the same purchase. What follows is illustrative token arithmetic, not a measured cost and not a bill ceiling. Assume one person, one month: 600,000 uncached input tokens and 60,000 output tokens. Rates come from the live Kunavo catalog, per million tokens.

One setting moves this more than the client choice does. LibreChat's titleModel pins which model writes conversation titles, and titles are high-volume, low-value traffic. Assume 200 new conversations a month at 1,200 input and 40 output tokens each: on Claude Sonnet 4.6 that background work estimates at $0.34, against $0.11 on Claude Haiku 4.5 — for output nobody reads closely. Clients without a title-model setting bill that traffic at whatever your main model costs.

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 and external tools sit outside this example, and the retrieval step in a RAG-capable client is not covered either — Kunavo serves no embedding model, so that runs locally or against an outside provider. The minimum top-up is $10 in prepaid credit, which is a funding minimum rather than a subscription or a task fee; see billing details.

Where to go next

If Open WebUI is your shortlist, LibreChat vs Open WebUI compares the two on setup friction specifically, and the Open WebUI setup page has the connection values. If documents are the job, AnythingLLM vs Open WebUI covers that pair. For desktop clients, Kunavo publishes setup pages for Cherry Studio and Chatbox; for roleplay, see SillyTavern and the API comparison for it. The protocol references are the OpenAI-compatible endpoint and the Anthropic base URL, which is what LibreChat's provider: anthropic route needs. Every one of those is published configuration, not a compatibility test. Create a Kunavo account when you have picked a client and want a key to test it with.

Choosing the endpoint rather than the client? OpenAI-compatible API and OpenRouter alternatives cover that decision, and AI cost optimization covers the method for measuring a real bill rather than estimating one.

FAQ

Repository status, stars, licences and push dates read from the GitHub API on September 19, 2026 and re-checked on September 21, 2026; LibreChat's compose file, example config and release list read the same way. Product prices and configuration behaviour come from each project's own documentation, read September 19 and re-checked September 21, 2026, and are not exhaustive — Cherry Studio's provider settings, Msty Nexus pricing and Open WebUI's enterprise price could not be read and are marked as such above. No client on this page was runtime-tested against Kunavo. Kunavo token rates come from the live catalog, and every dollar example here is illustrative token arithmetic.