Skip to main content
0 reactions

Rasa Tutorial — Part 1: What Is Rasa? Rasa Pro and Mantle Explained

Title card reading What Is Rasa?, Rasa Pro and its Mantle engine explained for complete beginners, with a chat window inside a loop labelled framework, LLM and tools.
Rasa Tutorial, Part 1: What Is Rasa?

Rasa lets you build AI agents that chat and talk with your customers and staff while following your business rules. This first part of the Rasa Tutorial explains what Rasa is, what Rasa Pro and its free Developer Edition are, and how Mantle, Rasa’s newest engine, runs a conversation.

By the end of this part you’ll know what the series builds, what it costs, and what to have ready before you build your first agent in Part 2.

This is Part 1 of the Rasa Tutorial. The parts in Module A, Start Here:

All parts of the Rasa Tutorial

Module A: Start Here

Module B: Building Skills

  • Part 3: Write a Skill’s Instructions
  • Part 4: Tools: Let Your Agent Run Python
  • Part 5: Memory: Remember What Matters
  • Part 6: References: Answer from Your Documents
  • Part 7: Responses: Control the Exact Wording

Module C: Progressive Control

  • Part 8: Tool Constraints
  • Part 9: Scoped Instructions
  • Part 10: Ordered Blocks
  • Part 11: Sub-skills and Composition

Module D: Orchestration

  • Part 12: The Runtime Loop: How Mantle Handles a Turn
  • Part 13: Guarantees and Guardrails
  • Part 14: Skills and Routing
  • Part 15: Context Management

Module E: Agent Setup

  • Part 16: agent.yml: Identity, Persona and Rules
  • Part 17: integrations.yml: Claude, Model Groups and Routing
  • Part 18: MCP Tools
  • Part 19: Hooks

Module F: Testing

  • Part 20: Simulate and Evaluate Your Agent
  • Part 21: Scenarios and Assertions
  • Part 22: Build Faster with the Rasa Copilot

Module G: Voice

  • Part 23: Talk to Your Agent in the Browser
  • Part 24: Designing for Voice
  • Part 25: Connect a Phone Line
  • Part 26: Test and Monitor Voice

Module H: Channels

  • Part 27: The REST Channel and a Website Chat Widget
  • Part 28: Microsoft Teams

Module I: Observability

  • Part 29: Tracing with Langfuse
  • Part 30: Metrics and Conversation Events

Module J: Deploy

  • Part 31: Run Your Agent with Docker
  • Part 32: CI/CD for a Rasa Agent
  • Part 33: Deploy to Kubernetes with Rasa’s Helm Chart
  • Part 34: Load Testing
  • Part 35: Deploy on AWS
  • Part 36: Deploy on Azure
  • Part 37: Deploy on Google Cloud

Module K: Rasa Studio

  • Part 38: Rasa Studio with a Mantle Agent

Module L: Wrap-up

  • Part 39: Final Project, Cheat Sheet and What to Learn Next

Cost: free. This part runs nothing.

Note: Mantle is in beta in Rasa Pro 3.21, so screens and files may change in later releases. This page is kept up to date.

What Is Rasa?

Rasa is a framework for building conversational AI agents: assistants that answer questions and complete tasks in a chat window, a messaging app or on the phone. Rasa calls it the code-first interface for building conversational AI agents (opens in a new tab) whose tasks must follow explicit business logic.

“Code-first” means you describe the agent in plain files: settings in YAML, instructions in Markdown and small helper functions in Python. You keep those files in version control, and your team can review and test them like any other code.

What makes Rasa different from a chatbot that sends everything to a large language model (LLM) is where the rules live. The LLM understands what the person means, but your business rules stay in your files, and the framework decides what is allowed to happen. Rasa runs on your own computer or servers, and you choose which LLM it uses.

Rasa Pro and the Free Developer Edition

Rasa Pro is the name of the software: the Python package rasa-pro. The newest release when this page was written is Rasa Pro 3.21.4 (opens in a new tab), from 12 November 2026.

Rasa Pro needs a licence key to run. The Developer Edition is Rasa’s free licence key for it: the same software, free to use within these limits, from Rasa’s Developer Terms (opens in a new tab) and pricing page (opens in a new tab):

WhatDeveloper Edition
PriceFree
Where you can use itYour own computer, a cloud development environment, Docker Desktop, and production within the limits below
ConversationsUp to 1,000 a month for an external use case, or 100 a month for an internal one
Use casesOne per company
Licence term12 months, and you can ask Rasa to renew it
TelemetryMust stay on
SupportThe community forum

Bigger deployments need Rasa’s Enterprise licence, which has no public price: you contact Rasa’s sales team. The older free package, Rasa Open Source (rasa), is in maintenance mode. Its last release, 3.6.21, came out in January 2025 and runs only on Python 3.10 and older, so this series uses Rasa Pro.

What Is Mantle?

Mantle (opens in a new tab) is the engine inside Rasa Pro that runs your agent’s conversations. Rasa calls it the orchestrator. You describe what the agent can do as skills. A skill is a folder with a skill.md file of plain-language instructions, plus, when it needs them, tools (Python functions), memory and exact wording. Mantle picks which skill is active, hands over between skills, calls your tools and keeps track of what it has learned.

The key idea is where control lives. Mantle runs one LLM in a loop, but the framework, not the model, is in charge of that loop. Rasa calls this deterministic-first: Mantle settles everything it can on its own before calling the model, and checks every model decision against the rules you set. You can start with plain instructions and add hard rules, called guarantees, only where a mistake would be expensive, such as “never send money without the customer’s confirmation”.

Mantle arrived as a beta feature in Rasa Pro 3.21, in October 2026. Rasa’s earlier engine, CALM, is still supported, but Rasa’s own FAQ (opens in a new tab) says it is no longer the recommended one for new agents. That’s why this series uses Mantle.

How Mantle Handles One Message

Every message a person sends runs through the same loop, described on Rasa’s Runtime Loop (opens in a new tab) page:

Diagram of how Mantle handles one message: the framework acts first, builds a focused prompt, the LLM decides, and tool calls go back through the framework before the loop repeats.
Every message runs the same loop, and the framework acts before the LLM.
  1. The framework acts first. It runs steps that are already decided, such as the next step of an ordered block (a fixed sequence of steps you define), and narrows which skills and tools the LLM may use.
  2. It builds a focused prompt with only the active skill’s instructions, the tools allowed right now and the memory that matters.
  3. The LLM decides. It replies, calls a tool, switches to another skill or waits, but only within what the framework allowed.
  4. Tool calls go back through the framework. Constraints and confirmations are checked before a tool runs, and then the loop starts again.

One message can go round the loop several times, for example to call a tool, read its result and then reply. Anything the framework settles on its own is faster and cheaper than a model call, and a person can’t talk the agent out of it.

What’s Inside a Mantle Agent

A Mantle agent is a folder of plain files. Rasa’s project structure reference (opens in a new tab) lists every file it recognises:

Diagram of a Mantle project folder: the required agent.yml, integrations.yml and a skills folder with a skill.md file, plus optional memory, responses, references and tools.
A Mantle agent is two settings files and a skills folder.
  • agent.yml holds the agent’s identity, persona and rules for every conversation.
  • integrations.yml says which LLM drives the agent and which channels people reach it through.
  • Each skill has its own folder under skills/, and its only required file is skill.md, the instructions in plain words.

Memory, exact responses, tools and reference documents are added only when a skill needs them. Rasa treats a folder as a Mantle project when it has integrations.yml and at least one skills/*/skill.md.

Advertisement

Which LLM This Series Uses

Mantle needs an LLM for the judgment step of its loop. It works with many providers, set in integrations.yml: OpenAI, Azure OpenAI, Anthropic, Google Gemini, self-hosted models that speak the OpenAI API (such as vLLM or Ollama), and any other provider the LiteLLM library supports.

This series uses the Claude API from Anthropic, with the model Claude Sonnet 5.5. Rasa’s integrations.yml reference (opens in a new tab) shows this setup: the provider anthropic and an ANTHROPIC_API_KEY.

Cost: the Claude API is pay-per-use, so you pay for what your agent sends and receives. Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, on Anthropic’s pricing page (opens in a new tab). Every part that calls the model says what it really cost.

Mantle also uses an embedding model to search your reference documents. Anthropic doesn’t offer one (opens in a new tab), so the part on references picks one for you.

What Mantle Can’t Do Yet

Mantle is in beta, and Rasa’s What’s Next (opens in a new tab) page lists features that are designed but not built yet. These are the ones that matter most for real agents:

  • Saving conversations to a database, such as PostgreSQL.
  • Searching an external knowledge store such as Qdrant or Milvus. Today Mantle searches only the reference files inside your project.
  • Running your tools on a separate server.

Each one gets its own part in this series as soon as Rasa ships it.

How This Series Works

The Rasa Tutorial covers every Mantle capability, in two streams that run side by side:

  • How-to parts. Each part teaches one concept with its own real, simple example, such as a skill that answers a shop’s opening hours. You can start at any part. The full list is in the “All parts” box above.
  • Product series. Each one builds a complete agent that solves a problem most organisations share, from an empty folder to a deployed product, with every step shown.

The product series build these 12 agents:

  • Company Knowledge Assistant: answers from the company’s own documents, with the source.
  • Employee Service Desk: one place for IT, HR, Finance and Facilities requests.
  • 24/7 Customer Service: resolves common requests and hands over to a person with the full conversation.
  • Approvals and Requests: raises requests, applies the rules and routes them to the right approver.
  • Complaints and Cases: records complaints and tracks them against their deadlines.
  • Appointments and Scheduling: books, moves and cancels appointments.
  • Sales Lead Qualification: answers website visitors, qualifies them and books a demo.
  • Billing and Payments: explains bills, sends payment links and sets up payment plans.
  • New-Hire Onboarding: guides new joiners through their first weeks.
  • Customer Onboarding: guides new customers through sign-up and verification.
  • Recruitment Screening: answers job questions, screens applicants and books interviews.
  • Supplier and Invoice Questions: tells suppliers where their invoices and payments stand.

What You’ll Need for Part 2

Common Mistakes

MistakeWhat to do instead
Installing the rasa packageThat’s Rasa Open Source, which is in maintenance mode. Install rasa-pro.
Thinking the Developer Edition is a separate productIt’s the free licence key for Rasa Pro, the same software.
Building several different agents on one free keyThe Developer Edition covers one use case per company. More need Rasa’s Enterprise licence.
Expecting Mantle to keep conversations after a restartSaving conversations to a database isn’t built yet. Keep long-running work, such as an open request, in your own systems.
Following CALM tutorials for a new agentRasa no longer recommends CALM for new agents. Build with Mantle.
Writing important business rules only as instructionsUse guarantees for anything a mistake would make expensive. The series covers them in Part 13.

FAQ

Is Rasa free?

Rasa Pro is free with the Developer Edition licence key, for one use case and up to 1,000 external or 100 internal conversations a month. The LLM you connect costs money separately.

Is Rasa open source?

The original Rasa Open Source is, under the Apache 2.0 licence, but it’s in maintenance mode. Rasa describes Rasa Pro as “paid and open code”: you can read its code, but you use it under Rasa’s licence terms.

Is Mantle ready for production?

Rasa marks Mantle as beta in Rasa Pro 3.21. The Developer Edition allows production use within its limits, but check the What’s Next list first: for example, conversations aren’t saved to a database yet.

Do I need to know Python?

No. Skills are plain-language instructions in Markdown. Tools are short Python functions, and the series explains every line the first time it uses one.

Can Rasa run on my own servers?

Yes. Rasa runs on your own infrastructure, and you choose a hosted LLM or one you run yourself.

Conclusion

Rasa Pro is a code-first framework for AI agents, and the Developer Edition makes it free within clear limits. Mantle, its newest engine, keeps the framework in charge: it settles what it can by itself, lets the LLM make only the judgment calls, and checks every decision against your rules. In the next part you’ll create your first Mantle agent and talk to it.

Next → Part 2: Create Your First Mantle Agent

Additional Resources

Rasa, Claude and other names mentioned are trademarks of their respective owners. JD Bots isn’t affiliated with or endorsed by Rasa Technologies or Anthropic.

0 comments