parlant

Parlant is an open-source interaction control harness for customer-facing AI agents. It addresses the problem of prompt overload and fragile routing graphs by dynamically narrowing conversational context on each turn, so only relevant rules, knowledge, and tools enter the LLM prompt. Core concepts include Guidelines (condition-action behavioral rules), Relationships (dependencies and exclusions between guidelines), Journeys (multi-turn SOPs), Canned Responses (pre-approved templates to prevent hallucination), Tools (activated only when matching observations), Glossary (domain vocabulary), and Explainability (OpenTelemetry tracing). The project is designed for teams building support, sales, onboarding, or advisory agents in regulated or high-stakes domains where tone, accuracy, compliance, and auditability matter. It is LLM-agnostic and can integrate with frameworks like LangGraph, LlamaIndex, or Agno by wrapping external workflows as tools. Parlant is distributed as a Python package (Python 3.10+) under the Apache 2.0 license, with an official React chat widget available separately.