21 Open-Source Chatbot Platforms Worth Building On in 2026

The chatbot stack has changed more in the past twelve months than in the previous five years. The closed frontier - GPT-5.5, Claude Opus 4.7, Gemini 3.1 Ultra - keeps pushing reasoning and tool use forward, while the open-weight side has gone from "interesting" to "production-ready" thanks to DeepSeek V4, GLM-5.1, Qwen 3.6, Kimi K2.6, MiniMax M2, and Xiaomi MiMo-V2. Wrapping any of those models into a real product still takes a platform: something to manage prompts, conversation state, retrieval, tool calls, channels, and observability.

Open-source chatbot platforms sit in that wrapper layer. They give you the most control, the lowest long-term cost, and - for regulated industries - the only way to keep traffic and data off third-party infrastructure. They also ask the most from you: hosting, scaling, security patches, model upgrades, prompt versioning, and the long tail of edge cases that show up only after real users hit the bot.

This guide walks through 21 open-source projects worth knowing, who they're for, and where each one shines.

When a managed agent platform makes more sense

Open source is the right call when you have engineering capacity and either strong customization needs or hard data-residency constraints. For most support, sales, and onboarding use cases, the math tips the other way. You spend weeks wiring channels and retrievers instead of weeks improving the bot's actual behavior.

Berrydesk is a managed alternative built specifically for customer support. You launch a branded AI agent in four steps: pick a model, train it on your sources, brand the widget, and deploy. Under the hood you can route across GPT-5.5 and GPT-5.5 Pro, Claude Opus 4.7 and Sonnet 4.6, Gemini 3.1 Ultra and Pro, plus open-weight options like DeepSeek V4 Flash, Kimi K2.6, GLM-5.1, Qwen 3.6, and MiniMax M2 - sending routine intents to a cheap fast model and reserving the frontier ones for hard escalations. Training sources include websites, PDFs, Notion, Google Drive, and YouTube transcripts. AI Actions handle bookings, refunds, order lookups, and payment flows. Channels include the web widget, Slack, Discord, WhatsApp, and more.

If you want speed-to-launch and the model picker without owning the infrastructure, Berrydesk is the shortcut. If you want to own every layer of the stack, the rest of this list is for you.

How to read this list

Three rough categories show up below. Conversational AI frameworks (Rasa, Botpress, Microsoft Bot Framework) come from the pre-LLM era and have since added LLM support. Modern LLM-app frameworks (LangChain, Haystack, Dify, Flowise, LangFlow) are built around prompts, retrieval, and tool calls from the start. Self-hosted chat UIs (LibreChat, Open WebUI, AnythingLLM, Chatwoot) ship a usable interface on day one and let you wire models behind it. Pick the layer your team actually needs - most production bots end up combining two of the three.

1. Rasa

Website: rasa.com

Rasa is still the reference open-source framework for intent-driven conversational AI, especially in regulated and on-premise contexts. The Python codebase, granular NLU pipeline, and rule-plus-ML dialogue policy give you control that pure LLM stacks rarely match. Rasa Pro layers in CALM (Conversational AI with Language Models), which lets you orchestrate frontier or open-weight LLMs while keeping deterministic guardrails on top. For finance and healthcare deployments where every reply needs to be traceable, Rasa remains hard to beat.

• Granular NLU control - train your own intents, entities, and pipelines instead of trusting a black-box LLM.
• Hybrid policies - combine ML, rules, and LLM-driven policies inside one dialogue manager.
• Audit-friendly - every transition is inspectable, which matters for compliance reviews.

2. Botpress

Website: botpress.com

Botpress sits between developer tooling and visual no-code, and that hybrid is its real strength. The visual flow builder is approachable for ops teams; the TypeScript modules are extensible for engineers. The platform has leaned hard into LLM agents and tool use over the past two years, with first-class connectors for OpenAI, Anthropic, and open-weight providers. If you want a self-hostable platform that doesn't force every flow change to go through a pull request, Botpress is one of the few that delivers on it.

• Visual flow builder with code escape hatches when flows get complex.
• Modular runtime - swap NLU, channels, and storage adapters without rewriting the bot.
• LLM-native - agents, tool calls, and knowledge bases are first-class, not bolted on.

3. Microsoft Bot Framework

Website: dev.botframework.com

The Microsoft Bot Framework is the enterprise default when you're already on Azure. Bot Framework SDK ships in C#, JavaScript, Python, and Java, and slots into Azure AI Foundry, Azure OpenAI, and the broader Microsoft 365 surface. The newer Copilot Studio sits on top and gives non-developers a low-code path to the same runtime. For Teams-first organizations that want to deploy across web, Teams, and voice with one codebase, the framework still has the deepest integration story of any option here.

• Azure-native - single sign-on, identity, and observability come for free.
• Multi-channel - web, Teams, Slack, SMS, voice, and more from one bot.
• Enterprise governance - fits cleanly into Microsoft 365 compliance and tenant policies.

4. Botkit

Website: github.com/howdyai/botkit

Botkit is a Node-first toolkit that pre-dates the Bot Framework's prominence and was eventually absorbed by Microsoft. It's still a good fit for small teams who want a thin, hackable layer over Slack, Webex, Teams, or Facebook Messenger without inheriting the full Bot Framework surface. Activity has slowed, but the codebase is mature and the patterns it established (middleware, conversation state, dialog management) influenced almost every framework that followed.

• Node.js native - easy to fit into an existing JavaScript or TypeScript backend.
• Lightweight - minimal abstraction over the underlying messaging APIs.
• Battle-tested patterns - middleware, dialogs, and conversation context done well.

5. OpenDialog

Website: opendialog.ai

OpenDialog targets the conversation design crowd, with a model-driven authoring interface and strong support for regulated industries. Healthcare, insurance, and government deployments are its sweet spot, partly because it makes conversation flows reviewable artifacts rather than scattered prompt strings. The platform integrates LLMs for understanding and generation, but the underlying behavior is governed by a structured conversation graph - useful when "the AI made it up" is not an acceptable answer to a regulator.

• Conversation-design-first - reviewable, versionable flow artifacts.
• Regulated-industry features - fine-grained audit, PII handling, on-prem deploy.
• LLM-augmented, not LLM-led - generative outputs are constrained by the graph.

6. Claudia Bot Builder

Website: npmjs.com/package/claudia-bot-builder

Claudia Bot Builder is a tiny library that strips away the boilerplate of deploying a serverless chatbot. It pairs naturally with AWS Lambda and lets you write the bot's logic as a single function while it handles webhook signatures, channel quirks, and message formatting for Messenger, Telegram, Slack, Skype, and Viber. It's not a fit for complex agentic workflows, but for a focused FAQ or lead-capture bot, it's refreshingly small.

• Serverless-friendly - built around AWS Lambda from day one.
• Multi-channel with one logic file.
• Minimal surface area - easy to read the whole library in an afternoon.

7. Tock

Website: doc.tock.ai

Tock is a French open-source platform that's been quietly running production bots for SNCF and other large European deployers for years. It supports text and voice, ships with its own NLU stack, and connects to dozens of channels including Google Assistant, Alexa, and WhatsApp. The platform has also added LLM connectors so you can mix intent-driven flows with LLM-generated responses where it makes sense.

• Voice and text in one platform - useful for IVR, kiosk, and accessibility use cases.
• European data residency - deployable entirely within EU infrastructure.
• Mature production track record in transportation and public services.

8. BotMan

Website: botman.io

BotMan is the de facto chatbot framework for the PHP ecosystem. If your team already runs Laravel or Symfony, BotMan lets you reuse your existing models, queues, and auth without standing up a separate Node or Python service. It supports the major messaging platforms and has a Laravel-flavored API that PHP developers will find immediately familiar. Adoption is smaller than the big Python and JavaScript frameworks, but for PHP shops it's the path of least resistance.

• PHP-native - integrates with Laravel, Symfony, and existing PHP stacks.
• Multi-platform - Slack, Telegram, Messenger, web, and more.
• Familiar idioms for PHP developers.

9. Bottender

Website: bottender.js.org

Bottender is a Node.js framework with a React-flavored, declarative API for chatbot logic. It works well for conversational UIs where you want to think in components and props instead of imperative dialog managers. The library is no longer the most actively developed option on this list, but the codebase is solid and the design choices age well.

• Declarative API - describe conversation behavior the way you'd describe React components.
• Multi-channel adapters for Messenger, LINE, Slack, Telegram, Viber, and Web.
• TypeScript support out of the box.

10. DeepPavlov

Website: deeppavlov.ai

DeepPavlov is a research-grade conversational AI library from MIPT in Moscow. It ships pre-trained models for intent classification, named entity recognition, slot filling, question answering, and dialogue. If your team has ML engineers and you want to fine-tune the underlying components rather than treat them as a black box, DeepPavlov gives you the dials. It also slots in nicely as the NLU layer behind a higher-level framework.

• Research-grade NLP components - pre-trained models you can fine-tune.
• Composable pipelines - build intent → NER → response in a single config.
• Strong for non-English - solid Russian and multilingual support.

11. The open-weight LLM tier (DeepSeek V4, Kimi K2.6, GLM-5.1, Qwen 3.6, MiniMax M2, MiMo-V2)

Websites: deepseek.com, moonshot.ai, z.ai, qwen.ai, minimax.io, github.com/XiaomiMiMo

This isn't a single platform but it's the most important "open-source chatbot" entry in 2026, so it earns its slot. DeepSeek V4 Flash runs at $0.14 per million input tokens with a 1M-token context window. Moonshot's Kimi K2.6 supports 12-hour autonomous coding sessions and swarms of up to 300 sub-agents. Z.ai's GLM-5.1 is MIT-licensed, scores 58.4 on SWE-Bench Pro (ahead of GPT-5.4 and Claude Opus 4.6 on that benchmark), and was trained entirely on Huawei Ascend chips. Alibaba's Qwen 3.6-27B is dense, Apache-licensed, and beats much larger MoE rivals on agentic coding. MiniMax M2 lands at roughly 8% the price of Claude Sonnet at 2x the speed. Xiaomi's MiMo-V2-Pro ships >1T total parameters with 42B active under MIT.

For a chatbot platform, what this means in practice is that the model layer is no longer the bottleneck. You pair these weights with one of the orchestration frameworks below - LangChain, Haystack, Dify, Flowise - and you have a frontier-class agent running on hardware you control.

• Cost collapse - routine support traffic at fractions of a cent per resolution.
• Long context - 1M tokens lets you stuff entire knowledge bases without elaborate RAG.
• MIT/Apache licensing on several variants makes air-gapped, on-prem deploys viable.

12. Wit.ai

Website: wit.ai

Meta-owned Wit.ai is still around and still free. It's an NLU service rather than a full chatbot platform - you send it text and get back intents and entities. For small projects that need a hosted understanding layer without standing up Rasa or DeepPavlov, it remains a low-friction option. The trade-off is the usual one with free Meta services: limited control over the roadmap and a quiet pace of feature releases.

• Free hosted NLU - no infrastructure to run.
• Multilingual - over a hundred languages supported.
• Simple REST API - easy to bolt onto any backend.

13. ChatterBot

Website: chatterbot.readthedocs.io

ChatterBot is a Python library that learns responses from example conversations using simple statistical matching. It's not state-of-the-art and it shouldn't be powering production support, but it's an excellent teaching tool. If you're learning how chatbots work or building a side project, it gets you to a working bot in a single afternoon.

• Beginner-friendly - minimal Python, clear concepts.
• No external API needed - runs entirely locally.
• Adaptive - learns from each conversation by default.

14. Typebot

Website: typebot.io

Typebot is an open-source visual builder for conversational forms and chatbots. Think Typeform crossed with a flow editor. It's self-hostable, has a clean drag-and-drop UI, and supports LLM blocks for generative responses inside otherwise structured flows. For lead capture, onboarding, and survey-style use cases, it's one of the fastest paths from idea to deployed bot.

• Visual flow builder with conditions, variables, and integrations.
• Self-hostable - full control over data.
• LLM blocks for generative responses inside structured flows.

15. Juji

Website: juji.io

Juji focuses on what it calls cognitive AI assistants - chatbots tuned for empathetic, interview-style conversations. It's used for HR screening, onboarding interviews, and qualitative research where you want the bot to read between the lines. The cognitive layer is proprietary, but Juji exposes enough customization that it earns a spot on most open-platform shortlists.

• Designed for empathetic, interview-style conversations.
• No-code authoring - non-developers can build and iterate.
• Personality and emotion modeling baked into the runtime.

16. LangChain

Website: langchain.com

LangChain is the most widely adopted framework for building LLM applications, full stop. It gives you abstractions for prompts, chains, agents, tools, retrievers, memory, and evaluation. The accompanying LangGraph library has become the standard way to build stateful multi-agent systems, and LangSmith handles observability. The ecosystem is sprawling - sometimes overwhelmingly so - but every model you'd want to use in 2026 has a LangChain integration on day one.

• Universal model coverage - OpenAI, Anthropic, Google, DeepSeek, Moonshot, Z.ai, Alibaba, MiniMax, plus local runtimes.
• LangGraph for stateful, multi-agent workflows.
• LangSmith for tracing, evaluation, and prompt versioning.

17. Haystack

Website: haystack.deepset.ai

Haystack from Deepset is a production-grade framework for retrieval-augmented generation and question answering. It's pickier than LangChain about its abstractions, which makes it slower to start with but cleaner to maintain at scale. If your chatbot lives or dies by document retrieval - long PDFs, technical manuals, knowledge bases - Haystack's pipeline model and pluggable document stores are worth the upfront investment.

• RAG-first design - retrievers, rankers, and readers as first-class components.
• Pluggable document stores - Elasticsearch, OpenSearch, Weaviate, Qdrant, pgvector, and more.
• Production-minded - strong evaluation and deployment story.

18. Dify

Website: dify.ai

Dify is one of the fastest-growing open-source LLM application platforms. It combines a visual workflow builder, prompt management, RAG pipelines, agent tools, and an observability layer in one self-hostable package. It supports the major closed and open-weight providers out of the box, including DeepSeek, Kimi, GLM, Qwen, and MiniMax alongside OpenAI, Anthropic, and Google. For teams that want a Berrydesk-style experience but want to host it themselves, Dify is the closest analog.

• Visual workflow builder with both chatflow and agent-style execution.
• Built-in RAG and prompt management.
• Wide provider support, including all the major open-weight families.

19. Flowise

Website: flowiseai.com

Flowise is a no-code visual builder for LangChain and LlamaIndex flows. You wire nodes on a canvas - model, prompt, retriever, tool, output parser - and Flowise generates the executable graph behind the scenes. It's a great way to prototype agent workflows that you'll later harden into code, and the self-hosted deployment is straightforward. For internal tools where the team wants to iterate on prompts without touching a repo, Flowise is hard to beat.

• Visual node editor for LangChain and LlamaIndex flows.
• Embeddable widgets and APIs for shipping flows into apps.
• Marketplace of pre-built templates to start from.

20. LibreChat

Website: librechat.ai

LibreChat is an open-source chat UI that mirrors the polish of ChatGPT but lets you bring your own models. It supports OpenAI, Anthropic, Google, and any OpenAI-compatible endpoint, which means you can plug in self-hosted DeepSeek V4, Kimi K2.6, or GLM-5.1 deployments. Multi-user authentication, conversation sharing, plugins, and assistants are all built in. It's a strong choice for an internal "company ChatGPT" without sending traffic to a vendor.

• Polished multi-user chat UI with auth and roles.
• OpenAI-compatible - plugs into nearly any model endpoint.
• Plugins, assistants, and code interpreter built in.

21. Open WebUI

Website: openwebui.com

Open WebUI started as a frontend for Ollama and has grown into a full self-hosted assistant platform with RAG, web search, document chat, model routing, and a plugin system. It's the easiest way to stand up a private ChatGPT-style interface in front of locally running open-weight models. For internal use cases on regulated networks - legal, defense, healthcare - it's quickly becoming the default.

• Best-in-class UI for local models via Ollama and OpenAI-compatible servers.
• RAG, web search, and document chat built in.
• Active community and rapid release cadence.

Bonus mentions worth knowing

The 21-platform list above hits the established players, but a few projects didn't quite fit the chatbot-platform framing yet are worth keeping on your shortlist.

AnythingLLM is a self-hostable RAG-first chat application that handles document ingestion, vector storage, and chat in one container. LangFlow is a visual builder similar to Flowise, originally LangChain-flavored, now provider-agnostic. Chainlit is the easiest way to put a polished chat UI in front of a Python LLM script - useful for internal tools and demos. Letta (formerly MemGPT) focuses on long-term agent memory, which becomes interesting when your bot needs to remember a customer across months of conversations. Hugging Face Chat UI is the open-source frontend behind HuggingChat and pairs naturally with TGI-served open-weight models. Chatwoot is the open-source customer support suite - not a chatbot framework itself, but the inbox most self-hosted bots end up routing to for human handoff.

How the model landscape changes the build-vs-buy math

Three shifts have reshaped what an open-source chatbot stack looks like compared to even a year ago.

The first is context. Claude Opus 4.6 and Sonnet 4.6 ship with a 1M-token window at no surcharge. Gemini 3.1 Ultra goes to 2M. DeepSeek V4 Flash and Kimi K2.6 both offer 1M. For a support agent, that means you can keep the entire knowledge base, the full conversation history, and the relevant policy documents in-context. RAG becomes a tuning lever for cost and latency - not a hard requirement for getting answers right.

The second is cost. DeepSeek V4 Flash at $0.14/$0.28 per million tokens, MiniMax M2 at roughly 8% the price of Claude Sonnet at 2x speed, and Qwen 3.6-27B running on a single GPU mean that routine support traffic now resolves at fractions of a cent per conversation. The economically interesting architecture is a routed one: send the easy 80% of intents to a cheap open-weight model, escalate the hard 20% to Claude Opus 4.7, GPT-5.5 Pro, or Gemini 3.1 Ultra.

The third is agentic tool use. Kimi K2.6 runs 12-hour autonomous coding sessions and coordinates 300 sub-agents across 4,000 steps. GLM-5.1 sustains an 8-hour plan-execute-test-fix loop. Claude Opus 4.7 leads SWE-Bench Pro at 64.3%. For a customer support agent that needs to actually book appointments, process refunds, and orchestrate multi-step recovery flows, these models flip AI Actions from demoware to dependable.

If you're building from scratch on open source, plan for these three shifts. Architect for routed multi-model inference, not single-model lock-in. Treat retrieval as one tool among several, not the whole architecture. And design your tool surface - bookings, refunds, lookups - assuming the model can actually use it well.

Common pitfalls when self-hosting

A few patterns show up over and over in teams that go open-source and regret it eighteen months later.

Underestimating ops. Hosting one model is fine. Hosting five providers, an embedding model, a reranker, a vector store, an inbox, and a dashboard - with auth, observability, backups, and failover - is a small engineering team's full-time job.

Skipping evaluation. Open-source frameworks make it easy to ship a bot. They don't make it easy to know if the bot is getting better or worse week over week. Plan for an eval harness from day one, or you'll be flying blind by month three.

Picking a framework before picking a use case. LangChain is overkill for a five-intent FAQ bot. Rasa is overkill for a single-LLM document chat. Match the framework to the actual problem.

Forgetting human handoff. The bot will get things wrong. The question is whether a real agent can take over the conversation in two seconds with full context, or in two minutes with none. Tools like Chatwoot exist precisely for this.

Closing the loop

Open-source chatbot platforms in 2026 are better than they've ever been, and the open-weight models behind them have closed most of the gap to the closed frontier. If you have the team and the appetite, building on Rasa, Botpress, Dify, or LangChain on top of DeepSeek V4 or GLM-5.1 will give you a powerful, owned stack for a fraction of the per-token cost of a vendor.

If you'd rather spend that engineering time on the bot's behavior - its tone, its knowledge, its actions - and let someone else handle the model routing, the channels, the widget, and the inbox, that's where Berrydesk comes in. Pick a model, train it on your sources, brand the widget, ship it to your site, Slack, Discord, or WhatsApp. Try it for free at berrydesk.com.