Skip to main content
This guide takes you from an empty folder to a working, tested agent that you can chat with or talk to. You’ll install Mantle, then use the Rasa Copilot, a set of MCP tools that plug into your IDE, to build, evaluate, and refine your first skill in a tight loop. The skill you build here works on both text and voice without changes. Mantle adapts how it speaks to the channel the conversation arrives on, so a voice agent is the same project with speech switched on.

The build loop

Agent development works in a loop. Rasa provides you with a harness that allows you to accelerate that loop and if you’re ready for it - run it on autopilot.

1. Install Mantle

Mantle ships in the rasa-pro package. Install it into a fresh virtual environment on Python 3.10 to 3.13:
Check it worked:
Mantle is currently pre-release, so pin the exact version rather than letting pip resolve the latest stable one.
You will also need a Rasa license key and an API key for your LLM provider. The setup wizard in the next step asks for both. To talk to your agent by voice, you also need an API key for a speech provider. This guide uses Deepgram. Voice is optional: everything in this guide works over text without it.

2. Create a project

Run rasa init to start the setup wizard. It also connects the coding agents you pick to the rasa-tools MCP server:
The wizard walks you through the full setup:
  1. Asks where to create your project
  2. Asks which coding agents to set up (Claude Code and Cursor are pre-selected; VS Code and OpenAI Codex are also available), then installs the Mantle skills for them, registers the rasa-tools MCP server, and downloads the Mantle documentation
  3. Collects your keys and writes them to a .env
  4. Validates and trains your agent
  5. Opens the Inspector so you can chat with it immediately
To pick the coding agents without the prompt, pass them with --ides, for example rasa init --engine mantle --ides claude,vscode. With --no-prompt and no --ides, Claude Code and Cursor are set up; pass --ides '' to skip coding-agent setup. Once it completes, your project looks like this:
That is what the scaffold writes. A project can hold more: a project-root memory.yml for cross-skill state, responses.yml for verbatim wording, references/ for knowledge, and tools/ for tools shared between skills. See Project structure for the full layout and which files are required.

3. Set your environment variables

The wizard collects your keys and writes a .env. To change them later, edit it directly:
.env
The wizard does not ask for DEEPGRAM_API_KEY. Add it yourself if you want to try voice in step 7, and set the Inspector to use Deepgram Flux in integrations.yml:
integrations.yml
This guide uses Deepgram, but you can use the speech providers of your choice: See Speech providers to switch. To use a different model, edit the group named by orchestrator.model_group in integrations.yml.

4. Connect your coding agent

If you picked your coding agents during rasa init, their Mantle skills are already installed, the rasa-tools MCP server is registered, and the full Mantle documentation is bundled at .rasa/docs/mantle/. Open the project in your coding agent and it has the context it needs to build and iterate on Mantle skills, plus the reference to look up anything the skills do not cover. For an existing project that wasn’t created with rasa init, if you skipped coding-agent setup, or to add another coding agent (including JetBrains IDEs), run the setup wizard in the project folder:
It detects the Mantle project and connects the IDEs you pick to the Rasa MCP server, so your coding agent can validate, train, and evaluate your agent. It also downloads the documentation and offers to install the skills. To add only the skills and the documentation, run:
Install the skills without the documentation and your agent still works, but it falls back to fetching the reference over the network — which many IDE sandboxes block.

5. Build your first skill from a prompt

With the Copilot connected, describe the skill in plain language, or paste in real transcripts of the conversations you want it to handle:
1

Describe it or paste transcripts

In your IDE chat, tell the Copilot what you want:
Or paste a handful of real support transcripts and ask the Copilot to turn them into a skill.
2

Let the Copilot scaffold the agent

The Copilot writes the skill.md, any tools/, and wires up config. Review the diff before accepting.
skills/check_balance/skill.md
Run rasa train, rasa run, and rasa inspect from the agent project root, the directory that contains agent.yml and integrations.yml. See Project structure.rasa train is the validation gate. Before packaging, it checks the whole project for undeclared memory writes, unknown response names, and malformed conditions. Treat its errors as authoring feedback and fix them before debugging conversation behaviour.

6. Evaluate with simulations

Rather than scripting every turn, let an LLM simulate a real user and score whether the agent met its goals. Ask the Copilot to generate and run a scenario:
The Copilot writes a scenario, validates it, runs the simulation against your running agent, and returns a pass/fail summary with a link to the full transcript. When something fails, it reads the result and recommends concrete fixes: a missing constraint, an ambiguous instruction, a tool that should be gated. The scenario format, the Mantle assertion checks, and how to read a result are in Simulate and evaluate.

7. Talk to your agent

Start your agent and open the Inspector in your browser:
You can test the same agent by typing or by speaking. Try both: the skill is identical, and Mantle changes how it responds to suit the channel.
Type a message such as Show me my recent transactions and step through each turn in the Inspector to see which skill was active, what memory was set, and which tools ran. Those moments are recorded as tracker events — see Tracker events.

8. Make it sound right on a call (optional)

Two small changes make the most difference on voice. Greet the caller. In a voice conversation, the caller expects the agent to speak first. The bundled session-start skill stays silent, so replace it with one that greets:
skills/builtin_session_start/skill.md
skills/builtin_session_start/responses.yml
Pick a voice. The voice is the model under language_map on the TTS model group you named from the inspector channel in step 3. Swap flux-alexis-en for any other Deepgram Flux voice, or adjust the pace with speed (0.5 to 1.5):
integrations.yml
To use a voice from another provider, set provider to azure, cartesia, or rime and add that provider’s API key to .env. See Speech providers. Run rasa train after the skill change, restart rasa inspect, and start a new voice conversation. For more on writing skills that work well out loud, see Designing for voice.

9. Loop

Feed what you learn back into the prompt: tighten the instructions, add a tool constraint, or add scoped instructions. Re-run the simulations, inspect again over chat and voice, and repeat until the agent behaves the way you need.

10. Take it to the phone

When the agent sounds right in the Inspector, connect it to a phone number through your telephony provider. The skills stay the same; you add a voice channel to integrations.yml. See Connect a phone line.

What’s next

Skills

Understand the building block you just created.

Instructions

Write the prose that drives your skill.

Tools

Give your skill actions it can take.

Progressive Control

Add guarantees where mistakes are expensive.

Voice agents

What Mantle does for you on a call, and what you control.

Build a voice agent

Take a card replacement skill from chat to a voice conversation.