AnythingLLM Review 2026: Free, Open Source and Private by Default, With a Security Record Worth Reading

Four ways to run it, one optional paid key, and a model engine you can swap.

The product comes in four forms that share a name but not a feature set. The free desktop app and the free Docker server are the full product; the hosted cloud and the Android app are narrower in different ways, covered below.

Swipe sideways to read the full diagram.

REAL_SCREENSHOT

REAL_SCREENSHOT

One of the larger open-source AI apps on GitHub, from a company that went through Y Combinator, with adoption numbers that mostly come from the vendor.

The docs say AnythingLLM is built by Mintplex Labs, Inc., founded by Timothy Carambat, and went through Y Combinator's Summer 2022 batch. The desktop privacy policy gives a Delaware corporation with an Anaheim, California address. The repository was created on 4 Jun 2023, is MIT licensed, and was last pushed to on 15 Oct 2026, the day before this review. No analyst market-size figure is quoted because no reliable public figure exists.

REAL_SCREENSHOT

REAL_SCREENSHOT

No Trustpilot page, a 2.7 Android rating, and a warm but older Hacker News launch thread.

There is no Trustpilot page for anythingllm.com, and no G2 or Capterra star score is quoted because none could be confirmed. There is no iPhone app: the mobile page offers only Google Play and a direct APK download. The only store rating is the Android app's, on the US Google Play listing on 16 Oct 2026: 2.7 stars from 237 reviews and a 50K+ download band. The vendor's mobile page says 68.6K+ downloads on Google Play; the store itself shows a lower band, and the difference is not explained.

REAL_SCREENSHOT

The three most visible written reviews are worth reading because the developer replied to two of them. Paraphrased, with dates and names as shown:

Hacker News has more, and mostly older, evidence. The Show HN thread of 5 Sep 2024 collected 368 points and 77 comments. Notable comments are below; two of the recent ones come from people promoting a rival or alternative project, which is noted on each card.

GitHub issues add a different kind of evidence: what the people who file and vote on requests still want. On 16 Oct 2026 GitHub search counted 306 open issues. The six most-reacted open requests are below, and the top one, fine-grained access controls, has been open since June 2024, which matters for teams that plan to share a server.

The issue tracker also shows how fast the project moves. The same search counted 3,708 closed issues on 16 Oct 2026, and from 16 Sep to 16 Oct 2026 it showed 177 issues opened and 168 closed; the counts move daily. The project's own “possible bug” label is on 11 open issues and 1,432 closed ones, and 256 open issues carry “feature request”. Labels are maintainers' triage, so this is not a defect rate. One open bug with eight reactions, opened on 12 Mar 2026, reports that switching to another thread while a reply is streaming can lose the final answer.

Independent write-ups exist, mostly from review sites with commercial links. RoboRhythms (1 May 2026, with an affiliate-commission disclosure, writing from its own testing) calls the desktop app one of the cleanest free chat-with-documents layers it has used, and argues that the word “local” misleads, because model inference needs a separate runtime such as Ollama or LM Studio. The app also ships its own built-in engine, per the vendor docs, so that second point is only partly true today. It also lists the lack of API rate limiting and of published retrieval-accuracy benchmarks as gaps. Its Cloud Basic limits (under 5 users, under 100 documents) are not on the vendor's Cloud page, so they are not repeated here.

Pattern: the praise is about low setup effort and a wide choice of model and vector-store back ends; the complaints are about the Android app's reliability, citation quality in 2024, resource weight and, in the written reviews, how much of the setup you still do yourself. About 226 Hacker News comments mention the name as of 16 Oct 2026, most of them in passing, so this is a sample, not a survey.

Free on your own computer or server; $50 or $99 a month if Mintplex hosts it; Enterprise by quote.

How the four forms differ in practice

The desktop app and the Docker server are free and run on your own hardware. Both can use the built-in model engine (on Docker it is optional and switched on with a setting), and both support MCP and custom agents. Cloud Basic at $50 a month and Cloud Pro at $99 a month run as a Mintplex-hosted instance on AWS with no built-in model, no MCP and no custom agents. The Android app is free and runs small on-device models; the sources do not say whether it supports agents. The docs name the cloud tiers Starter and Professional while the pricing page calls them Basic and Pro.

Sources: anythingllm.com/cloud, the cloud limitations page, the Docker quickstart and the mobile page, as of 16 Oct 2026.

REAL_SCREENSHOT

A worked example: a team of four, three ways

Suppose four colleagues want one shared place to ask questions of their own documents.

Route 1, a Docker server you run. The licence is $0. The docs list t3.small on Amazon Web Services, e2-standard-2 on Google Cloud and B2ps v2 on Azure as the minimum instance sizes, and offer one-click templates for Railway and Render. Those providers bill you directly, and their prices are not included here. One independent review site, RoboRhythms, suggests a small server of about $10 a month; that is its own estimate and is not confirmed here.

Route 2, Cloud Basic. $50 x 12 = $600 a year, with no yearly discount shown on the pricing page. Cloud Pro is $99 x 12 = $1,188, which is $588 a year more than Basic. In both cases the model bill comes on top, because the plans include no model.

How big is that model bill? It depends on the provider you connect. Take an illustrative month of 5 million input tokens and 1 million output tokens (a volume picked for this example, not a measurement). Using each vendor's own pricing page on 16 Oct 2026, per million tokens:

• OpenAI gpt-6-luna, $0.10 in and $0.50 out: 5 x $0.10 + 1 x $0.50 = $1.00. OpenAI gpt-6.1-sol, $2.00 in and $10.00 out: 5 x $2.00 + 1 x $10.00 = $20.00 (short-context rates).
• Anthropic Claude Haiku 5.5, $0.10 in and $0.50 out: $1.00. Claude Sonnet 5.5, $2 in and $10 out: $20.00 (Haiku rates apply to prompts up to 100,000 tokens).
• Open-model price scale, from Ollama's pricing page: gpt-oss:20b at $0.07 in and $0.30 out gives 5 x $0.07 + 1 x $0.30 = $0.65; gpt-oss:120b at $0.15 and $0.60 gives $1.35; kimi-k3 at $3.00 and $15.00 gives $30.00. Whether AnythingLLM connects to that hosted service was not tested, so read these three only as a scale for open-model prices.

Over a year on Cloud Basic, $600 plus $12 for a $1.00 month is $612, and $600 plus $240 for a $20.00 month is $840. The hosting fee dominates with a cheap model; a mid-priced model at this volume adds about 40% on top.

Sources: anythingllm.com/cloud, the Docker cloud-deployment docs, platform.openai.com/docs/pricing, docs.anthropic.com pricing, ollama.com/pricing and roborhythms.com, as of 16 Oct 2026. Result (our calculation); hardware, server and electricity costs are not included.

Not an affiliate link. It opens AnythingLLM's official website.

Model costs, an optional Pro key with no published price, and hosted limits that the pricing cards do not show.

The cloud cards say “Just bring an LLM API Key”. That means the model bill comes from OpenAI, Anthropic or whichever provider you connect, on top of the plan. The limits below come from AnythingLLM's own docs.

Where each limit comes from

Hosted cloud. The hosted instance has no GPU and limited CPU and RAM, so it ships no built-in model: you connect a model you host yourself or a cloud provider. MCP and custom agents are not supported there, for security reasons. The built-in embedder cannot take arbitrarily large files, and the docs warn that a 5,000-page PDF can crash the instance with a 502 error, so use an external embedder for big files.

Your own machine. For local models the docs ask for 16GB of RAM and an 8-core CPU for a basic experience, and call 8 to 12GB or more of GPU memory great. The Docker server needs at least 2GB of RAM, a 2-core CPU with AVX2 and 5GB of storage; AVX2 is a hard requirement with the default vector database, so old CPUs and some virtual machines need a different one. Hardware is the real cost of going local.

Pro key and web search. The optional Pro key removes the daily limits on Magic Echo, Beacon and Tab, but no price or free daily size appears on the docs pages. Web search defaults to You.com, with a daily per-network rate limit that you can lift by adding your own key, which also means search queries leave your device by default.

Sources: anythingllm.com/cloud, docs.anythingllm.com (cloud limitations, system requirements, Pro pages) and the Sep 2026 You.com blog post, as of 16 Oct 2026.

REAL_SCREENSHOT

REAL_SCREENSHOT

A prorated refund at the vendor's discretion, and cancelling deletes the instance for good.

Only the hosted cloud is a subscription with these terms. The text below is from the Cloud Terms and Conditions page, which states it was last updated on 1 Jul 2024.

Payment and outages. Fees apply, and you are expected to watch for failed payments through my.mintplexlabs.com or your contact email. If an instance goes offline, restarting it from the dashboard or contacting support is your responsibility, and there are no refunds or credits for downtime.

Refunds. In the terms' words, “If you request a refund, we will refund your subscription prorated for the time you have used the service during the current billing cycle.” No refunds are given for any other reason, at the sole discretion of Mintplex Labs. In practice you get unused days back, nothing more.

Cancelling and deletion. You can cancel any time through the dashboard or by email, and no lock-in period is described. But “Deletion of services is final and non-recoverable.” After cancelling, non-payment or termination, no information about your instance or data is retained, so export what you need first.

Vendor rights. Mintplex Labs may cancel a subscription or suspend access for any reason, at any time. Read that before putting a team's only copy of its work on the service.

docs.anythingllm.com Cloud Terms and Conditions, as of 16 Oct 2026. The Cloud privacy policy, last updated 1 Jul 2024, says data is stored on Amazon Web Services.

Start with the free desktop app, move to Docker for a team, and use the cloud only if you will not host anything.

Workspaces hold your files, chats and agent tools, and you choose the model, embedder and vector store behind them.

The docs describe AnythingLLM as an all-in-one app for RAG, AI agents and more without code or infrastructure work. You add documents to a workspace, they are split and embedded into a vector store (LanceDB by default; the docs also cover Chroma, Pinecone, Qdrant, Weaviate, Milvus, Zilliz and Astra DB), and chats retrieve and cite passages. The docs' LLM setup list names six local options (including LM Studio, Ollama and KoboldCPP) and 15 cloud providers, including Anthropic, OpenAI, Google Gemini, AWS Bedrock, Mistral, Groq and OpenRouter. The vendor's own comparison pages claim 40+ providers; the docs list is the smaller verified number.

The recent releases show where the work is going. Version 1.17.0 added per-chat reasoning-effort controls for several providers and Hugging Face model deep links, and rebuilt the Meeting Assistant. Version 1.18.0 added reasoning controls for local and cloud models and a domain allowlist for embed widgets. The Meeting Assistant runs a speech model on your device, with no bot joining the call, but its summaries use whichever model you configured, so a cloud model means the transcript text goes to that provider.

REAL_SCREENSHOT

REAL_SCREENSHOT

REAL_SCREENSHOT

REAL_SCREENSHOT

REAL_SCREENSHOT

Install, pick a model, add documents, then decide whether web search and agents should be on.

A worked example from the vendor docs: why one question works and another does not

Say you drop a folder of recipe PDFs into a workspace. The docs describe the path in order. Uploading turns each file into text. Moving the file into the workspace splits that text into chunks, and each chunk becomes a vector stored in the vector database. When you ask a question, it is converted the same way, the nearest chunks are found, and chunks scoring under the similarity threshold (20% by default) are dropped. Only 4 to 6 chunks go to the model as context.

That explains a limit the docs spell out. A question such as “get me the title of the 3 high-calorie meals” across hundreds of recipes gets refused, because no handful of chunks can answer a question about the whole collection; the docs say that kind of job belongs to agents. A question whose answer sits inside one passage is where plain document chat works.

If answers come back empty or wrong, the docs suggest this order. First use Query mode, which answers only from your documents and says so when it finds nothing. If it reports no relevant information, set the similarity threshold to No restriction, or reword the question with the terms your document uses. Next try Accuracy Optimized search, which re-ranks more chunks and, per the vendor's own testing, adds about 100 to 500 milliseconds; it is available only with the default LanceDB store. As a last resort, pin a document so its full text goes into the context window, which gives better answers at the cost of speed and money and only works if the document fits.

One privacy detail from the same pages: embedding a document makes it available to every thread in the workspace, and in multi-user mode to every user who can open that workspace.

docs.anythingllm.com chat modes, Attaching vs RAG and the older RAG page (which the docs mark as legacy; its recipe example is still the clearest statement of the limit), last updated 9 Sep 2026, read 16 Oct 2026.

REAL_SCREENSHOT

REAL_SCREENSHOT

Agent skills, no-code flows, scheduled jobs and MCP work on Desktop and Docker; the hosted cloud gets none of the extensible parts.

The docs list built-in agent skills for RAG search, web browsing and scraping, saving files, summarizing documents, charts, a SQL agent, a file-system agent, document generation, and Gmail, Google Calendar and Outlook agents. Agent Flows is a no-code builder with blocks such as web scraper, LLM instruction and file read and write. Scheduled jobs run an agent on a cron schedule. MCP servers are added through a configuration file or the in-app manager on Desktop and Docker, and the docs say it works with MCP tools generally.

Skills and flows. Built-in agent skills and your own custom skills (written with a developer guide) work on Desktop and Docker; custom agents are not supported on the hosted cloud. Agent Flows, the no-code chains of blocks, are also Desktop and Docker. MCP. Tools from MCP servers reach agents on Desktop and Docker, not on the hosted cloud. Scheduled jobs. Recurring agent runs with a cron builder are documented in the Desktop docs.

The docs also list a developer API, a browser extension, a Telegram channel and a Desktop Assistant overlay for talking to your agent; they are not covered in depth here.

docs.anythingllm.com (AI agents, MCP compatibility, scheduled jobs, cloud limitations), as of 16 Oct 2026.

REAL_SCREENSHOT

Download and savings figures that only the vendor can check.

The desktop privacy policy, effective 14 Jul 2025, says “over 1M people have downloaded the app”, and the home page gives no date for the 7M+ figure, so the two numbers cannot be compared directly. The savings and token counters have no stated method, and local tokens that were never billed cannot be audited from outside. Treat them as marketing.

25 published advisories in the repository, most fixed; local-first by design, but web search and usage telemetry are on by default.

GitHub lists 25 published security advisories for the repository on 16 Oct 2026: 1 critical, 7 high, 11 medium and 6 low, 23 of them published in 2026. They are published by the project itself and carry fixes for most, which is a sign of an active disclosure process. It is also a lot of findings for a product that may hold a team's private documents, and several are about multi-user isolation and embed widgets. A few list no fixed version, so check the advisory for your version.

REAL_SCREENSHOT

What leaves your device, and what does not

The app itself. The Desktop privacy policy says none of your messages, chat histories and documents are ever transmitted from your system and that everything is saved locally by default. It also says some anonymous usage data is collected, names PostHog as the telemetry service, and says you can opt out in the app settings.

Defaults that reach out. You.com is the default search provider on Desktop, Mobile and the server, and the vendor says its endpoint retains nothing. A cloud model provider you choose receives your text under that provider's policy, so use a local model for sensitive material. The Pro key is validated periodically against a vendor licence server, and more than 48 hours offline invalidates it; air-gapped licences are available by contacting the vendor.

Hosted cloud. Data is stored on Amazon Web Services and deleted on cancellation, under a policy last updated on 1 Jul 2024. Ask for current terms if you are regulated.

Licence. The Desktop Terms of Use (version of 14 Jul 2025) grant a non-exclusive, non-transferable licence for personal and internal business purposes, bar modifying the Software or creating derivative works, and cap aggregate liability at $50. The GitHub source is MIT licensed, so for redistribution or white-labelling use the MIT source or ask the vendor.

Android app. Google Play data safety says no data is shared with third parties; it may collect app activity and app info and performance, and data cannot be deleted. Read the listing before installing.

docs.anythingllm.com privacy, cloud and Pro pages, the Sep 2026 You.com blog post and the US Google Play listing, as of 16 Oct 2026. JAVIS is not a law firm; read the full Desktop terms before relying on this summary.

REAL_SCREENSHOT

People and small teams who want private AI on hardware they control and can tolerate some tinkering.

A narrow cloud, a long advisory list, an unpriced Pro key and thin independent evidence.

Other self-hosted chat front ends and local model apps.

How they differ. Open WebUI has about 2.3 times AnythingLLM's GitHub stars (155,821 against 67,161) but changed its licence in April 2025, which matters if you plan to rebrand it for more than 50 users. LM Studio now sells hosted open-model plans at $20 and $100 a month alongside a free local app, and its home page is now titled LM Studio Bionic. AnythingLLM's hosted plans host the workspace and leave models to you. Jan lists 44,961 stars and LibreChat 45,763, and LibreChat's home banner says it is joining ClickHouse.

Ollama, which AnythingLLM stopped bundling in v1.18.0, has its own JAVIS review. Its pricing page shows that running models on your own machine is free and that paid plans are for its hosted cloud models. For jobs beyond what Agent Flows cover, n8n sells by workflow execution: its Starter plan lists 2.5k executions and Pro 10k. If you outgrow the default vector store, the docs list Chroma as an option, and LlamaIndex is the developer-side route when you would rather build the retrieval pipeline in code than use a finished app. AnythingLLM's own comparison pages are vendor material; competitor claims there count only where the competitor's own page confirms them.

REAL_SCREENSHOT

Comparing self-hosted AI and agent tools? These reviews cover the nearest ground.

Each card opens a published JAVIS review.

AN ALTERNATIVE

Open WebUI review 2026

Another self-hosted AI front end, and its licence change.

LOCAL MODELS

Ollama review 2026

The local model runner AnythingLLM stopped bundling in v1.18.0.

MCP

Claude MCP connectors 2026

How MCP connectors work, useful before you add MCP tools to AnythingLLM.

DICTATION

Wispr Flow review 2026

A dedicated dictation app, for comparison with Magic Echo.

VECTOR DATABASE

Qdrant review 2026

A vector database AnythingLLM can use instead of its default store.

AUTOMATION

n8n review 2026

Workflow automation for jobs beyond what agent flows cover.

BUILD IT IN CODE

LlamaIndex review 2026

The developer route when you want to build the retrieval pipeline yourself.

VECTOR STORE

Chroma review 2026

One of the vector databases the AnythingLLM docs list as an option.

MODELS

Hugging Face Hub review 2026

Where many of the open models you would run locally are published.

Questions people actually ask before installing

Is AnythingLLM free?

Yes. The desktop app and the Docker server are free; the code is MIT licensed per GitHub. Optional paid items are the hosted cloud (Basic $50 and Pro $99 a month per anythingllm.com/cloud on 16 Oct 2026) and a Pro key for desktop Magic features, whose price is not published on the docs pages.

Does it run offline?

Yes, with the built-in engine or another local model, and the vendor says the desktop app can run entirely offline. Web search and any cloud model need a connection.

What does the cloud plan include?

A private hosted instance with a custom subdomain on Basic, priority resources and a 72-hour support SLA on Pro, with your own LLM API key. It has no built-in model, no MCP and no custom agents, per the vendor's limitations page.

Does it support MCP?

Yes, on Desktop and Docker. Not on the hosted cloud.

Is it open source?

The code on GitHub is MIT licensed. The downloadable desktop app also carries Terms of Use (version 14 Jul 2025) that limit it to personal and internal business use and bar derivative works, so redistribution needs care.

Is it secure?

GitHub lists 25 published advisories, 1 critical and 7 high, most with fixes. Keep the app updated, restrict access to a server, and read advisories that match your version.

Is there an iPhone app?

The mobile page offers Android only, through Google Play or a direct APK. The US Google Play listing shows 2.7 stars from 237 reviews.

Can I get a refund on the cloud?

The terms promise a refund prorated for the current billing cycle if you request it, at Mintplex Labs' discretion, and cancelling deletes the instance permanently.

What do users say?

The 2024 Show HN launch thread drew 368 points and mostly positive comments about easy setup, with some bug reports and feature requests; 2026 comments describe daily use with local models. A May 2026 review site praises the desktop app but warns that “local” still needs a separate model runtime. There is no Trustpilot page, and the Android app is rated 2.7 on Google Play.

Is AnythingLLM an affiliate product here?

No. This review has no affiliate links.

Where this came from, and when.