LobeChat vs LibreChat — the rename, the acquisition, and which one to self-host

LobeChat and LibreChat are both alive, and both changed identity after most comparisons of them were written: LobeChat is now LobeHub under a source-available licence, and LibreChat is now a ClickHouse project that is still MIT. The short answer: choose LibreChat when several people share one deployment and every endpoint, model list and permission should live in version-controlled configuration; choose LobeHub when one person or a small team wants a polished agent workspace with a desktop app, and will accept a heavier stack and a licence with a derivative-work condition.

Two traps sit on this query. On the LobeHub side it is the rename — search results, forks and tutorials still describe 1.x, and 2.0 deleted the authentication and storage variables those tutorials teach while keeping the provider ones, so a stale walkthrough half works. On the LibreChat side it is the acquisition, which turned a formerly paid add-on into a free self-hosted service, so older write-ups quote a subscription that has been announced for retirement.

What each project actually is today

Repository figures read from the LobeHub and LibreChat GitHub APIs on September 19, 2026. Do not read either version from the GitHub Releases tab: LobeHub's releases are nightly pull-request builds and desktop canary prereleases, and LibreChat flags every release as a prerelease so /releases/latest returns nothing. Both projects' package.json is the honest source, and whether v0.8.8 has shipped stable was still unresolved on the day of this check.

The rename is documented rather than inferred: maintainer arvinxx wrote in Discussion #10007 on November 3, 2025 that "our product name will be upgraded from 'LobeChat' to 'LobeHub'". It is also incomplete — the LICENSE is titled LobeHub Community License while its body still calls the product LobeChat. The acquisition is equally on the record: ClickHouse's announcement says founder Danny Avila and his team joined, and that for existing deployments "operations continue unchanged".

Which one to pick, before the feature table

Two of those rows deserve the detail. LibreChat's access-control documentation describes three layers — per-role feature permissions, per-resource access-control lists, and system grants such as manage:users — with the built-in ADMIN and USER roles always present. LobeHub's workspace permissions are three fixed roles, Owner, Member and Viewer, where configuring providers, setting member budgets and viewing the audit log are all Owner-only and a Viewer cannot chat at all. Neither model is strictly richer for every team: LibreChat gives you finer grain and more to get wrong; LobeHub gives you one chokepoint and less to configure.

Migration cost is the other thing to price before you choose. Neither project documents an importer for the other's conversation history, and the stores have nothing in common — LibreChat runs on MongoDB, while LobeHub 2.0 runs only on a server-side PostgreSQL database. Treat a switch as a fresh deployment plus a manual move of prompts and agent definitions.

Agents, tools and knowledge bases

Both have moved well past "a chat window over an API", and both moved in 2026, so a comparison written from 2025 material is describing neither.

Dates are worth carrying here. LibreChat's v0.8.6 added Agent Skills and subagents, and v0.8.7 added native Anthropic custom endpoints, bundled the Admin Panel and released the Code Interpreter as open source — the changelog dates that release June 23, 2026 while the GitHub API stamps it June 24, so pick one and say which. LobeHub's public changelog is patchier: the newest file in its changelog directory is dated August 3, 2026, against a repository pushed on September 18, so roughly six weeks of shipped changes are not described in any user-facing changelog.

One boundary that applies to both: each indexes knowledge-base documents as vector entries — LobeHub's pricing page meters them by the entry — and that indexing step needs an embedding model, which Kunavo does not serve. Run it against a local model or an outside provider, whichever client you choose; it is a separate account from the chat key in both setups.

What it costs to run either one

Sources checked September 19, 2026: LobeHub's self-hosting overview, Docker Compose page and Cloud Sandbox variables; LibreChat's Docker install and Code Interpreter pages. Two operational footnotes worth having before you start: LibreChat's docs warn that the default MongoDB image needs AVX and crashes on Apple Silicon unless you pin mongo:4.4.18 in an override file, and LobeHub's compose page warns that an existing deployment upgraded without the new gateway settings will restart the gateway container in a loop and fall back to running agents in the browser, printing an ACTION REQUIRED warning in the app container's logs.

One thing this page could not settle: what a self-hosted LobeHub's default sandbox provider actually sends off your own machine. The documentation states the default (SANDBOX_PROVIDER=market), that MARKET_BASE_URL points at https://market.lobehub.com unless you change it, that an Onlyboxes-compatible provider can replace it, and how a trusted-client credential pair is configured — but not what content the default sends to that service. If that matters to you, read the setting before you enable the built-in code, shell and file tools.

LobeHub Cloud prices, and the number the page shows twice

LobeHub's plan cards default to the yearly toggle, so a quick look reports $9.9, $19.9 and $39.9. The plan comparison table below them on the same page gives the monthly figures. Both are true — $118.8 ÷ 12 = $9.90 — but only one of them is what you pay to be billed monthly.

Read from lobehub.com/pricing on September 19, 2026. Three qualifications. The Free tier is model-gated rather than merely small: in the comparison table fourteen frontier model rows, GPT-6 Astra among them, read "Not supported" in the Free column, which is why the hero line offering a free trial of GLM-5.3-Flash and GPT-6 Astra sits directly above a Free card that marks GPT-6 Astra unsupported — confirm in the product, not from the headline. The three paid tiers say additional credit packages are available for purchase without stating a size or price anywhere on the page. And the Enterprise bullet "Self-hosted Provider" is never defined on that page, so this page does not say what it includes; it is not a requirement to self-host LobeHub, which the Community License already permits.

Bring-your-own provider keys are not a paid feature: that row carries a checkmark in all four columns including Free, read from the icon markup rather than from a summary. Whether LobeHub Cloud accepts an arbitrary custom base URL, as opposed to your own key for one of its listed providers, is not stated in the documentation and this page does not claim it either way. The self-hosted path is the one the documentation covers.

What a LobeHub credit is worth, by our arithmetic

LobeHub publishes credits per million tokens per model but never a credit-to-dollar rate, and the pricing FAQ that might answer it renders in the browser rather than in the page source. So the conversion below is our arithmetic, not LobeHub's statement. On the five rows we could check against a vendor's standing list rate, their credit figure equals that rate exactly, which implies 1,000,000 credits ≈ US$1.00 of list-price usage. The sixth row, Gemini 3.8 Flash, matches that vendor's current introductory rate rather than its standing list price — the same arithmetic against a temporarily lower number, and a reminder that the pattern tracks whatever a vendor charges this month.

Credit figures from LobeHub's credit table, list prices from each vendor's published rate as recorded in Kunavo's own catalog, both on September 19, 2026. If the inference holds, Premium's 15,000,000 credits are about $15 of list-price model usage for $24.9 billed monthly, and Free's 500,000 credits are about $0.50 a month. Treat that as a reading of a pattern, not a quoted rate — LobeHub could reprice any row without changing anything it has published, and the page states no conversion you could hold it to.

Pointing either client at one API key

This is the sharpest practical difference between them, and it is about doors rather than features. LibreChat's documented door is a file. LobeHub's documented doors are environment variables and a settings pane, and neither is a file you can review in a pull request.

LibreChat's custom endpoint takes both protocols. Leaving provider unset gives you the OpenAI-compatible client; setting provider: 'anthropic' routes the same block through the native Messages client, and the documentation says in so many words to use it "for Anthropic itself or Anthropic-compatible gateways that speak the native Messages API".

version: 1.3.16          # librechat.example.yaml, checked 2026-09-19
endpoints:
  custom:
    # OpenAI-compatible route — the default when no provider is set.
    - name: 'Kunavo'
      apiKey: '${KUNAVO_API_KEY}'
      baseURL: 'https://api.kunavo.com/v1'
      models:
        default: ['claude-sonnet-4-6', 'gpt-5-6-terra', 'gemini-3-8-flash']
        fetch: true            # fills the picker from GET /v1/models
      titleConvo: true
      titleModel: 'claude-haiku-4-5'
      modelDisplayLabel: 'Kunavo'

    # Native Messages route — same key, different protocol, no /v1 suffix.
    - name: 'Kunavo (Messages)'
      provider: 'anthropic'
      apiKey: '${KUNAVO_API_KEY}'
      baseURL: 'https://api.kunavo.com'
      headers:
        anthropic-version: '2023-06-01'
      models:
        default: ['claude-sonnet-4-6', 'claude-opus-5']
        fetch: false           # docs: models.fetch is not used here
      modelDisplayLabel: 'Kunavo'

Three limits to know before you rely on it. Dynamic model fetching is not used for native Anthropic endpoints, so that list is hand-maintained. If you set baseURL: 'user_provided' so each person supplies their own URL, LibreChat does not forward your configured header templates to that destination — header-based routing quietly stops working in that mode. And tokenConfig, the per-model rates LibreChat uses for context-usage display, cost breakdowns, balance transactions and multi-endpoint agent billing, is metadata you type yourself; it never learns a gateway's real rates, so both the figures shown and the credits deducted are only as correct as what you entered.

Self-hosted LobeHub keeps the provider environment variables its 1.x documentation taught — the authentication and storage ones are a different story, covered below. The one that bites is the /v1 suffix: LobeHub forwards to the URL you give it, and its own docs warn that a missing or extra suffix shows up as an empty assistant message rather than an error.

# LobeHub, self-hosted. The /v1 suffix belongs in the VALUE — LobeHub
# forwards to the URL you give it instead of rebuilding the OpenAI path.
OPENAI_API_KEY=sk-kn-...
OPENAI_PROXY_URL=https://api.kunavo.com/v1
OPENAI_MODEL_LIST=-all,+claude-sonnet-4-6,+gpt-5-6-terra,+claude-haiku-4-5

# Native Messages route, if you prefer it. Root only, no /v1.
ANTHROPIC_API_KEY=sk-kn-...
ANTHROPIC_PROXY_URL=https://api.kunavo.com

# 2.0 requires these two and no longer reads ACCESS_CODE, Clerk or NextAuth.
AUTH_SECRET=...
JWKS_KEY=...

Variables checked against LobeHub's .env.example and the model-provider page on September 19, 2026. The OPENAI_MODEL_LIST syntax is documented on the model list page: -all hides the built-in catalogue first, +id adds one back, and =Display Name renames it. In the UI the equivalent path is Settings then Language Model, and current source also ships a create-a-provider flow whose SDK type list includes openai, azure, azureai, anthropic, google, cloudflare, ollama and seven more.

One variable set you will be offered by search engines and should not use: CUSTOM_PROVIDER_COUNT, CUSTOM_PROVIDER_N_BASE_URL and CUSTOM_AI_PROVIDERS. They come from a feature request opened in February 2026 that maintainer arvinxx closed on March 20, 2026 with "we don't plan to support this feature in the community edition", and they appear nowhere in .env.example. Also check the age of any LobeHub walkthrough before following it: 2.0's breaking changes removed ACCESS_CODE, Clerk, NextAuth and the client-side database mode, and added AUTH_SECRET and JWKS_KEY as required. Kunavo's own LobeHub setup guide is filed under the former name and was written against the 1.x environment: the three provider variables it teaches are still correct in 2.x, but its surrounding assumptions about authentication are not.

A worked estimate for the model bill

The client is free either way, so the only line that scales is tokens. These figures are illustrative token arithmetic, not measured costs and not a bill ceiling. Assume one person's month of interactive chat: 400 turns, each sending 10,000 input tokens of accumulated context and receiving 600 output tokens — 4M input and 240K output tokens in total, with no caching. Rates are live Kunavo catalog prices per million tokens.

Read it against the shape of a subscription rather than as a winner. Under these assumptions that month costs $2.08 on Claude Haiku 4.5 and $10.40 on Claude Opus 5 — a spread wide enough that which model you pick matters more than which client you run it in. The same month priced in LobeHub Cloud credits at their published rates would be 5.2M credits on Haiku 4.5 and 15.6M on Sonnet 4.6, so this Sonnet month already runs past Premium's 15,000,000-credit allowance and into credit packages whose price is not published. That comparison is arithmetic on two published tables, not a like-for-like quote: a subscription also buys storage, sync and the hosted product, and metered credit buys none of that.

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. The minimum top-up is $10 in prepaid credit — a funding minimum, not a task fee or a monthly plan. See billing details.

Setting it up, and what to check first

Whichever you choose, the first request is the cheap test. Call GET /v1/models with your key before you debug a container: if it returns a list, the URL and key are fine and an empty picker is the client's model list, not the endpoint. For LobeHub, start at the LobeHub integration guide and read it as a configuration reference rather than a compatibility test — neither client was runtime-tested against Kunavo for this page. For LibreChat there is no Kunavo setup page yet; the YAML above plus LibreChat vs Open WebUI covers the same ground, and the client directory records what is and is not known about its endpoint behaviour. Create a Kunavo account when you are ready to fund a key.

If you are still choosing an interface rather than a provider, AnythingLLM vs Open WebUI compares two more self-hosted options, and the OpenAI-compatible API guide explains what any of these clients is actually sending.

FAQ

Checked September 19, 2026: both GitHub repository APIs, both package.json files, LobeHub's LICENSE and .env.example, the redirect from the old lobe-chat path, lobehub.com/pricing's own markup, HTTP status of librechat.ai/pricing and /enterprise, and LibreChat's custom-endpoint documentation source; re-checked September 21, 2026 against ClickHouse's acquisition announcement, LobeHub issue #12288 and both projects' self-hosting docs. Statements about LobeHub's self-hosting stack, permissions and sandbox, and about LibreChat's access control, balance system and Code Interpreter, come from each project's current documentation rather than from a deployment run here. Neither client was runtime-tested against Kunavo. Kunavo token rates come from the live catalog, and every dollar example is illustrative token arithmetic.