Skip to main content
These features are designed but not yet built. Some already appear in integrations.yml as configuration that parses and validates without a runtime consumer; those are called out below.

Responses and questions

Variant selection by channel or condition. Multiple variants are already parsed; only the first is used today. ask_<memory_entry> hooks for dynamically generated questions, so a skill can compute the wording of a collect: prompt at runtime.

Integrations

MCP servers. mcp_servers: already validates in integrations.yml, but there is no resolver. mcp/<server-id>:<tool-name> in import_tools does not work. Connecting external tool servers is planned. Knowledge stores. knowledge_stores: (Qdrant, Milvus, external FAISS) validates but is never queried. Today search_knowledge reads only the local index built from references/**/*.md. Per-skill retrieval scoping, so a skill’s own references/ are searched ahead of, or instead of, the shared index. Tracker store (Postgres, etc.) for persistent conversation state, and event broker (Kafka, etc.) for streaming conversation events.

Tool control flow

ToolResult.next is defined but not executed: returning RunTool(...) or ActivateSkill(...) is currently a no-op. Executing them is planned.

Remote tool hosting

rasa run tools will spin up a dedicated tool server that the runtime connects to over MCP, letting you host tools separately.

Agent-to-agent

A2A skill exposure will let you expose a skill so other agents can invoke it over the A2A protocol.