AnythingLLM vs LibreChat: Knowledge Workspaces or Multi-Model Agents

AnythingLLM and LibreChat are both free, MIT-licensed and self-hostable, so the choice is not price — it is whether retrieval is built into the product or deployed beside it. AnythingLLM makes documents the unit of work: drop files into a workspace and the embedder and vector store are already there. LibreChat makes the conversation the unit of work: many models, many users, per-role permissions, and document search as a separate service you stand up yourself.

One framing correction before the comparison, because it changes how you read the project. LibreChat is now a ClickHouse project: ClickHouse, Inc. announced the acquisition on November 4, 2025, saying that for existing deployments "operations continue unchanged with ongoing investment in the open-source project and community", and librechat.ai still carries the banner today. AnythingLLM remains a Mintplex Labs product. Every source cited below carries a September 21, 2026 check date; the closing note names the ones re-read directly for this page.

Which one fits the task you actually have

What each one costs, and what "Pro" means

Prices read from anythingllm.com/cloud and the pages linked above on September 21, 2026. Two traps sit in that table. First, "AnythingLLM Pro" names two unrelated things — the $99/monthly hosted Cloud tier, and a Desktop Pro licence that requires Desktop v1.15.0 or later and whose price is on no reachable official page — so a price quoted without saying which one is meaningless. Second, the vendor's own vocabulary disagrees with itself: the pricing page says Basic and Pro, while the Cloud limitations page says "Starter" and "Professional". Nothing first-party maps one set onto the other, so confirm at checkout rather than assuming they are the same tiers.

There is also a timing conflict worth knowing on the LibreChat side: a February 12, 2026 GitHub discussion quotes the plan to "evolve this offering and discontinue this API in its current form", targeting May 1, 2026, yet the pricing page on September 21, 2026 still tells existing subscribers they can manage their subscriptions. Whether that date was executed is not something this page can confirm.

The retrieval split, which is the real decision

AnythingLLM. Documents go into a workspace and the app indexes them itself. The server env file ships VECTOR_DB="lancedb" as the only uncommented value, with Chroma, Pinecone, Astra, pgvector, Weaviate, Qdrant, Milvus and Zilliz commented beside it. The catch is scope: the embedder documentation says "Embedding models are set system-wide and cannot be configured atomically per-workspace like LLMs can", and that changing the choice later means you "will have to delete your uploaded documents and re-embed them". That is the migration cost of this product, and it is paid in re-indexing time rather than money.

LibreChat. Document search is not in the chat app at all. The RAG API "runs as a separate FastAPI service backed by a PostgreSQL + pgvector database", configured through RAG_API_URL and EMBEDDINGS_PROVIDER (openai, azure, huggingface, huggingfacetei or ollama), with chunking defaults of 1500 and an overlap of 100. Agents reach it as the file_search capability, and an admin can withhold that per role. More to deploy; more to control.

Kunavo does not serve embeddings, speech-to-text or text-to-speech — no enabled model carries those endpoints — so in either product the embedder slot, and AnythingLLM's separate transcription and speech slots, point at a provider that does, while the chat slot is the one a Kunavo key covers. RAG implementation covers how to split those stages.

Agents, MCP and who is allowed to change what

Two cautions on that table. The LibreChat permission list tracks the 0.8.8-rc documentation, and since the project publishes no stable release tag, an older pinned image will not have all of it — check against the version you actually run. And the AnythingLLM role names come from its Security and Access page, which opens with "DOCKER VERSION ONLY! These settings are only available in the Docker version of AnythingLLM" — a sentence about where the settings live, not a statement that hosted Cloud is single-user, since the Cloud docs describe that as a multi-user deployment.

Pointing either one at a single key

Both accept an OpenAI-compatible endpoint, which is why running them side by side on one key is a fair way to decide. The configurations differ in ways that matter for a gateway.

# AnythingLLM — LLM Preference -> provider "OpenAI (generic)".
# Self-host env names, from server/.env.example on master:
LLM_PROVIDER='generic-openai'
GENERIC_OPEN_AI_BASE_PATH='https://api.kunavo.com/v1'
GENERIC_OPEN_AI_API_KEY=sk-kn-...
GENERIC_OPEN_AI_MODEL_PREF='claude-sonnet-5'      # typed by hand; no model-list fetch is documented
GENERIC_OPEN_AI_MODEL_TOKEN_LIMIT=1000000
# Embeddings are a SEPARATE provider slot with its own key and its own base
# path (EMBEDDING_ENGINE / EMBEDDING_BASE_PATH / EMBEDDING_MODEL_PREF), and so
# are the transcription and speech slots. A working chat endpoint proves
# nothing about any of them.

# LibreChat — librechat.yaml in the project root, beside your .env.
endpoints:
  custom:
    - name: "Kunavo"                      # required
      apiKey: "${KUNAVO_API_KEY}"        # required; resolved from .env
      baseURL: "https://api.kunavo.com/v1"  # required
      models:                             # required
        default: ["claude-sonnet-5", "claude-haiku-4-5"]
        fetch: true
      # tokenConfig is what the docs say supplies a custom endpoint's rates and
      # context windows — read by context usage, cost breakdowns and balance
      # transactions. prompt/completion are USD per million tokens; context is
      # that model's own window. All three are required, per model.
      tokenConfig:
        claude-sonnet-5:
          prompt: 2
          completion: 10
          context: 1000000

AnythingLLM's own documentation describes that provider as a "developer-focused llm provider - you should not use it unless you know what you are doing" that "may not function as intended if you input any configuration setting incorrectly". It does not state whether tool calling, agents or streaming survive the route, so this page does not claim either outcome. LibreChat's custom endpoint reference requires name, apiKey, baseURL and models, and adds the option AnythingLLM has no documented counterpart for: provider: anthropic routes the same block through the native Anthropic client, at the cost of listing model ids by hand, because "OpenAI-style models.fetch is not used for native Anthropic custom endpoints". One more gate if you let each user bring their own endpoint: with baseURL: "user_provided", LibreChat does not forward configured header templates to that destination.

Cost visibility is asymmetric too. LibreChat documents tokenConfig as what defines a custom endpoint's per-million rates and context windows, and as the input to its context usage, cost breakdowns and balance transactions — so an admin hand-writes it per model, as above; AnythingLLM has no documented per-model rate table of its own. Neither product should be treated as your ledger. Setup for both is published: AnythingLLM and LibreChat.

A worked monthly estimate

This is illustrative token arithmetic, not a measured cost and not a bill ceiling. Assume one workspace answering 600 questions a month, each sending 12,000 input tokens of system prompt, retrieved passages and short history, and returning 600 output tokens. Rates are live Kunavo catalog prices per million tokens. The right-hand column doubles the retrieved context to show the lever that matters here — how much you retrieve per question is a pricing decision, and it is identical in both products.

Read it against the subscription line rather than instead of it. Under these assumptions Claude Haiku 4.5 costs $3.60 a month in tokens, which sits well below AnythingLLM Cloud's $50/monthly Basic tier — but the two are not substitutes, because that tier buys a managed private instance and still requires you to bring a key on top. A self-hosted LibreChat or AnythingLLM Docker instance moves that $50 onto your own server bill instead.

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 subscription. See billing details, or create a Kunavo account when you have picked a client. For a deeper spend breakdown on either side, see AnythingLLM API cost and LibreChat API cost.

Weighing a third option? AnythingLLM vs Open WebUI and LibreChat vs Open WebUI run the same comparison against a shared alternative.

FAQ

Sources checked September 21, 2026. Re-read directly while writing this page: anythingllm.com/cloud, the AnythingLLM Cloud limitations, Desktop Pro, agent, MCP, Security and Access and embedder-configuration docs, the AnythingLLM server env example, LibreChat's custom-endpoint, interface, MCP and RAG API references, the LibreChat releases listing and package.json, code.librechat.ai/pricing, the HTTP status of librechat.ai/pricing and /enterprise, the Code Interpreter discussion, and the ClickHouse acquisition and ClickHouse Agents beta posts. Kunavo token rates come from the live catalog, every dollar example is illustrative token arithmetic, and Kunavo has runtime-tested neither client — the linked setup pages are configuration references, not compatibility tests.