Best Open-Source AI Customer Support Chatbots (2026) and Hosting Costs

For some teams, open source is the first filter on a support chatbot. They want to read the code, keep conversations on their own servers and avoid per-seat or per-resolution bills. This guide is for the developer, technical founder or IT lead making that call: it covers ten projects that come up when people ask for an open-source AI chatbot for customer support, sorts them into inbox-first and builder-first tools, and prices a self-hosted setup at three sizes.

Every licence below was read from the project's own LICENSE file on 19 October 2026, and every release date and star count comes from the project's GitHub repository on the same day. Server and model prices were read that day from DigitalOcean's and OpenAI's pricing pages. GitHub stars measure attention rather than quality, so treat them as a dated popularity signal.

PepoChat, which publishes this guide, is a hosted AI support agent and is not open source. It appears only where the comparison needs a hosted price and in the section on when open source is the wrong choice.

Which open-source chatbots are actually open source?

Six of the ten projects use licences the Open Source Initiative approves, two are open core and two are source-available. The Open Source Initiative's Open Source Definition requires, among other things, that a licence must not stop anyone from using the program in a particular field of endeavour. In our reading, licences that forbid uses such as a competing service or multi-tenant hosting do not meet that test.

• Open source: AnythingLLM and LibreChat (MIT), RAGFlow and Rasa Open Source (Apache 2.0), Tiledesk (MIT on the components we checked) and Botpress v12 (AGPL-3.0).
• Open core, meaning an open-source core with paid features in a separately licensed folder: Chatwoot and Flowise.
• Source-available, meaning the code is public but the licence restricts some uses: Dify and Typebot.

Source-available is a licence category, not a verdict on quality. Typebot's licence is the Functional Source License: you may do anything except compete with Typebot, for example by selling access to your instance, and each release converts to Apache 2.0 two years after it ships. Dify's licence is Apache 2.0 plus two conditions: no multi-tenant service without written permission, where one workspace counts as one tenant, and no removing the Dify logo or copyright from its frontend. Neither licence appears on the OSI's list of approved licences.

Licences and editions change, so read the LICENSE file on the day you deploy, not the README's description. Chatwoot's AI sits in its paid folder, Botpress's current platform is sold as Botpress Cloud with no self-hosted edition on its pricing page, and Flowise stopped development in 2026.

Inbox-first or builder-first: which kind do you need?

An inbox-first tool is a live chat and helpdesk where people answer customers, with AI added on top; a builder-first tool is a platform for building the bot itself, usually without a screen where a person takes over. Support needs both halves, because the AI should answer routine questions and a person should pick up the rest with the transcript, as our human handoff guide explains.

Most builder-first projects answer with retrieval-augmented generation (RAG): the bot searches your documents first and answers only from the passages it finds. The RAG explainer walks through the pattern without code. The table shows which of the five support-focused projects covers which half.

The common open-source pairing is a builder in front of an inbox. Typebot has a Chatwoot block that opens a Chatwoot chat so the visitor can reach a person, and Chatwoot's Agent Bot API, part of its MIT-licensed core, lets any bot you run answer first and hand the conversation to an agent. Tiledesk is the one project here that ships both halves in a single deployment.

The best open-source AI support chatbots compared (October 2026)

Of the ten projects in this comparison, four combine an OSI-approved licence with AI answers from your documents in the free edition: Tiledesk, AnythingLLM, RAGFlow and LibreChat. Tiledesk and AnythingLLM also document a chat widget for website visitors.

Licences, releases and star counts were read from each project's repository on 19 October 2026, and requirements from each project's documentation or README the same day. Botpress's stars are for its main repository; its last release is the v12 line.

Chatwoot: the open-source inbox, with AI on a paid licence

Chatwoot is the most popular open-source support inbox in this comparison, with 37,959 GitHub stars, and its AI is not in the free edition. The project puts website live chat, email, WhatsApp, Messenger and Instagram in one queue and shipped version 4.18.0 on 18 September 2026. Its LICENSE file puts everything under MIT except the enterprise/ folder, which needs a paid Chatwoot licence for production use and holds the Captain AI code.

Chatwoot's help centre states that the Community Edition does not include Captain. Self-hosted, Captain comes with Premium Support at $19 an agent a month or Enterprise at $99, plus your own OpenAI API key, according to the article on enabling Captain, updated 14 October 2026. The free route to AI is the Agent Bot API in the MIT core, which lets a bot you build and host answer first.

Chatwoot's self-hosting guide asks for 2 CPU cores, 4 GB of RAM and 20 GB of SSD as a minimum, or 4 cores, 8 GB and 50 GB for production, plus PostgreSQL 12 or later and Redis 6 or later. The Chatwoot alternative page compares its hosted plans with PepoChat.

Tiledesk: live chat, a bot builder and RAG in one deployment

Tiledesk is the closest open-source match to a hosted AI support agent: a chat widget, an agent dashboard, a visual bot builder called Design Studio and human handoff ship together. Its Docker Compose guide says the Community version answers from your own content with RAG, using Qdrant as the vector store and your key for OpenAI, Gemini, Anthropic, Groq or Cohere, or an Ollama server for local models.

Tiledesk's drawbacks are weight and some uncertainty. The compose file, which the guide labels beta, starts 15 containers, including MongoDB, Redis, RabbitMQ and Qdrant, and the README points production users to Kubernetes with Helm. The repositories we checked are MIT-licensed, but the server package's metadata lists AGPL-3.0, and the server's GitHub repository did not load on 19 October 2026, although its npm package was updated on 14 October. Tiledesk's own cloud starts at €149 a month.

Dify: a full RAG toolkit under a modified licence

Dify is the most-starred project in this comparison, and its Knowledge feature covers the whole pipeline from document ingestion to retrieval. Sources can be files, Notion or a website; you review the chunks, choose vector, full-text or hybrid retrieval, and publish the app on your site as a chat bubble or an iframe.

Dify's documented minimum is 2 CPU cores and 4 GiB of RAM, and its Docker Compose stack runs PostgreSQL, Redis, Weaviate, Nginx and a code sandbox. Website import needs a Jina or Firecrawl API key, and Dify's customer service tutorial notes a 50-page cap on website sync. Two limits matter for support: there is no inbox where a person takes over, and the multi-tenant clause rules out an agency running one instance with a workspace per client unless Dify agrees in writing.

AnythingLLM: MIT-licensed document answers with a website widget

AnythingLLM is the lightest MIT-licensed way to put answers from your documents on a website. A workspace holds your files, and the embed widget exposes that workspace as a chat bubble through a script tag or an iframe, in the Docker version only.

AnythingLLM's documented minimum is 2 GB of RAM, a 2-core CPU with AVX2 and 5 GB of storage, with the LanceDB vector database built in, so no separate database server is needed. It can also run a local model inside its container, which keeps conversations off third-party APIs. The embed's README makes abuse your responsibility, so set the per-embed and per-session chat limits before you publish. There is no inbox for a person to take over.

Typebot: a fair-source flow builder that hands off to Chatwoot

Typebot is a visual builder for guided conversations, such as lead forms and triage menus, embedded on your site as a standard box, a popup or a bubble. Its self-hosting documentation calls the project fair source and lists what the licence forbids, such as selling access to your instance. Self-hosting needs PostgreSQL and two Next.js apps, a builder and a viewer.

Typebot has no knowledge base of its own. Its OpenAI block can search vector stores kept in your OpenAI account, which OpenAI bills at $2.50 per 1,000 file-search calls, with storage beyond the first free gigabyte at $0.10 per GB a day. Its Chatwoot block opens a Chatwoot chat so the visitor can reach a person, which makes the two a natural pair.

Botpress and Rasa: the old open-source answers have moved on

Botpress and Rasa both began as open-source chatbot frameworks, and neither is a self-hosted AI support bot today. The botpress/botpress repository is MIT-licensed, but it holds the SDK, command-line tools and integrations for Botpress Cloud, not the platform. Botpress v12, the last self-hostable version, is AGPL-3.0 and lists NLU, an FAQ module and human handoff but no LLM answers. As of 19 October 2026, Botpress Cloud's Free plan includes 100 conversations a month with no top-ups; the Botpress vs Voiceflow comparison covers the hosted product.

Rasa Open Source says in its README that it is in maintenance mode, and its last release came out on 14 January 2025. Rasa 3.6 requires Python 3.8 to 3.10, and the Python release calendar lists all three as end of life, 3.10 since 1 October 2026. Rasa's current product is a platform whose free Developer Edition allows one bot and up to 1,000 external conversations a month, under a free Rasa licence rather than an open-source one.

LibreChat, RAGFlow and Flowise: three you will see on other lists

Three more projects are popular on GitHub but fit website support less well than their stars suggest. LibreChat is a self-hosted ChatGPT-style workspace with agents and file search behind OAuth2, LDAP or email sign-in: useful for your support team's internal answers, not built for anonymous visitors.

RAGFlow is a document-parsing RAG engine with APIs, and its README recommends starting with 4 CPU cores, 16 GB of RAM and 50 GB of disk. Flowise, a visual agent builder, stopped feature development on 29 July 2026 and was archived on 13 August 2026, according to the maintainers' Future of Flowise post. Flowise's Apache 2.0 code remains available to fork, but a new deployment would be yours alone to maintain.

How much does it cost to self-host a support chatbot?

A self-hosted support chatbot costs about $30 a month in server and model fees at 2,000 AI replies, about $86 at 10,000 and about $190 at 50,000, at list prices read on 19 October 2026. None of those totals include your time, which is usually the larger cost.

The assumptions, so you can swap in your own: DigitalOcean Basic Droplets from its Droplet pricing page, with weekly backups at 20% of the Droplet price; managed PostgreSQL and Spaces object storage from the medium size up; and gpt-4o-mini at $0.15 per million input tokens and $0.60 per million output tokens from OpenAI's API pricing page. We assume about 4,000 input and 300 output tokens per reply, so 1,000 replies use 4 million input tokens ($0.60) and 300,000 output tokens ($0.18), or $0.78. Embeddings with text-embedding-3-small, at $0.02 per million tokens, add cents.

Prices read on 19 October 2026 from DigitalOcean's Droplet, Managed Databases and Spaces pricing pages, OpenAI's API pricing page and Chatwoot's self-hosted plans page.

Three lines move the bill more than the server. The model comes first: Chatwoot's self-hosted Captain defaults to gpt-5.2, at $1.75 and $14 per million tokens, so 10,000 replies cost about $112 instead of $7.80 unless you change the model. The licence comes second: three agents on Chatwoot's Premium Support pay $57 a month, the same as three Startups seats on Chatwoot Cloud in our pricing index, so self-hosting Chatwoot's AI buys control rather than savings.

Time is the third and largest line: upgrades, restore tests, monitoring and patches. Our build vs buy guide prices developer days at the US median wage; running an existing project takes fewer hours than building one, but never zero. For comparison, PepoChat, our hosted product, costs $49 a month for 2,000 replies with nothing to run, $129 for 10,000 on the same Starter plan with overage at $10 per 1,000, and $529 at 50,000 replies, well above the self-hosted bill.

What security and compliance work do you take on when you self-host?

Self-hosting moves every control a vendor would run onto your team: patches, backups, access, abuse limits and data requests. None of it is exotic, but none of it stops. The list below is the minimum for a public support bot, and the AI chatbot security guide covers the threats in more depth.

• Patch on a schedule. Watch each project's releases and update the application, the operating system and every container image, including the database and queue images you never think about.
• Test restores, not only backups. A backup you have never restored is a hope rather than a control.
• Limit the public endpoint. A website widget is an unauthenticated door that spends your model budget, so set rate limits, such as AnythingLLM's per-embed and per-session caps, and a monthly spending limit with your model provider.
• Check telemetry defaults. AnythingLLM sends anonymous usage data unless you set DISABLE_TELEMETRY to true.
• Know where the text goes. Self-hosting the app does not keep conversations away from your model provider; only a local model does.
• Own data requests. Retention, export and deletion are yours to configure, as our GDPR guide for chatbots explains.

A hosted tool runs those controls for you, but you take its word for them. PepoChat lists what it does on its security overview, such as encrypted credentials and rate limits per visitor session, and does not claim SOC 2, ISO certification or a hosting region, which is one reason some teams self-host.

When is open source the wrong choice for a support chatbot?

Open source is the wrong choice when nobody on your team will own the server, or when the AI you want sits in a paid edition anyway. Open source is the right choice when control of the code and data, or very high volume, matters more than hours.

Self-hosting is the right call in four situations. Conversations must stay on infrastructure you control, and with a local model in AnythingLLM or Tiledesk they never leave it. You need to change the code or integrate deeply. You already run Docker and PostgreSQL in production, so the extra hours are small. Or your volume is high: at 50,000 replies a month, the self-hosted bill above is about a third of what PepoChat would charge, and Chatwoot adds WhatsApp, email and social channels in one inbox for the price of a server.

Self-hosting is the wrong call in four others. Nobody owns it, and a support bot that breaks on a Saturday with no one on call is worse than none. The AI is paid anyway: Chatwoot's Captain costs $19 an agent self-hosted, and Botpress's current platform is a cloud service. The licence does not fit: an agency hosting clients on one Dify instance needs Dify's permission, and reselling a Typebot instance is a competing use. Or the project ends: Flowise is archived and Rasa Open Source is in maintenance mode, so you would inherit the fork. Our support stack for a bootstrapped SaaS shows what a small team can run instead.

A hosted AI support agent makes the opposite trade. PepoChat puts every feature on the free plan, including 200 AI replies a month, a crawl of up to 200 pages, the team inbox and human handoff, but it is not open source: you cannot read its code, self-host it or choose its model, which is OpenAI's.

What to do next

Shortlist two projects from the table and run each on a $24 Droplet for a week against your real help pages and twenty real customer questions, noting the hours you spend. If the hours are the problem, compare plans on the pricing page or create a free PepoChat workspace and put the same questions to it. If you are tempted to build your own bot instead, the build vs buy guide prices that route.