← Back to The State of Conversational AI, Q3 2026

The Periodic Table of Conversational AI: Full Capability Matrix (2026)

As of August 2026. Capability cells verified against current official documentation. Y = native, P = partial (add-on / paid tier / product-not-API), N = not documented.

Matrix

Self-hosted assistant UIs

System RAG Memory Multi-user Files Tools/MCP Proactive Auth/tenant Deploy Licence
Open WebUI Y Y Y Y Y Y Y Y Open WebUI Licence (source-available, branding clause)
LibreChat Y Y Y Y Y N Y Y MIT
Lobe Chat Y Y Y Y Y P Y Y LobeHub Community Licence (verify)
AnythingLLM Y Y Y Y Y Y Y Y MIT

Frameworks / builders

System RAG Memory Multi-user Files Tools/MCP Proactive Auth/tenant Deploy Licence
LangChain Y Y N Y Y N N Y MIT (library)
LlamaIndex Y Y N Y Y N N Y MIT (library)
Dify Y Y Y Y Y Y Y Y Dify OSS Licence (Apache-2.0 + conditions)
Flowise Y P P Y Y N P Y Apache-2.0 core; Enterprise features commercial
Rasa Y Y N Y Y P N Y Apache-2.0 (OSS); RAG in Rasa Pro

CX / chatbot platforms

System RAG Memory Multi-user Files Tools/MCP Proactive Auth/tenant Deploy Licence / hosting
Chatwoot Y Y Y Y P Y Y Y MIT (CE) + proprietary EE; AI (Captain) is premium
Botpress Y Y Y Y Y Y Y P Current Cloud = SaaS; self-host only legacy v12
Voiceflow Y Y Y Y Y N Y N Proprietary, cloud-only
Sendbird Y Y Y Y P Y Y N Proprietary, cloud-only

Commercial cloud (developer APIs + agent platforms)

System RAG Memory Multi-user Files Tools/MCP Proactive Auth/tenant Deploy Licence
OpenAI (Responses/Assistants API) Y P N Y Y P N N Proprietary
Anthropic (Claude API) P Y N Y Y P N N Proprietary
Microsoft Copilot Studio Y P Y Y Y Y Y N Proprietary
Google Vertex AI / Dialogflow CX Y P Y Y Y Y Y P Proprietary (on-prem via Google Distributed Cloud)

What the verified data shows

  1. MCP / tool use is now table stakes. Almost every system scores Y on Tools/MCP; the only non-Y are Chatwoot (native tool-calling but MCP is community-only) and Sendbird (tools/integrations but MCP unverified). A capability that did not exist before late 2024 is now near-universal - the single clearest "standardised in 2026" signal, and it corroborates the GitHub/search-demand story elsewhere in the report.

  2. Proactivity is the real differentiator. Scheduled/triggered, agent-initiated action splits the field cleanly: present in Open WebUI, AnythingLLM, Dify, Chatwoot, Botpress, Sendbird, Copilot Studio, and Google; absent or partial in LibreChat, Flowise, Voiceflow, LangChain, LlamaIndex, Rasa, and the OpenAI/Anthropic APIs. This is where "chatbot" becomes "agent", and where products visibly differ.

  3. Persistent memory is common but still fuzzy. Many systems now claim memory, but a genuine persistent long-term store is often a newer add-on, a paid tier, or a product feature not exposed at the API level (hence the cluster of P in the commercial-cloud row). It is maturing, not settled.

  4. Deploy/control is the cleanest fault line. Open-source self-hostable on one side (the assistant UIs, the frameworks, Dify, and Chatwoot) versus cloud-only on the other (OpenAI, Anthropic, Voiceflow, Sendbird, Copilot Studio; Google partial via Distributed Cloud; and notably Botpress, whose current product is now cloud-only). Control and convenience remain a real trade, not a solved problem.

  5. No single system is strong on control + conversation + agency at once. The self-hosted tools own control (deploy/auth/multi-user); the commercial platforms own managed polish and proactivity; and memory sits unfinished across both. That unresolved middle is the category's current frontier.

Key sources

← Back to The State of Conversational AI, Q3 2026