Skip to content

Download and use Agent Studio

Nuvoca Agent Studio contains five skills for creating and reviewing your voice Agent configuration. The package helps with system prompts, agentic routing, tools, knowledge search and knowledge files.

Download Agent Studio 0.1.9 (ZIP)

Version: 0.1.9 · Package: manual installation · Size: about 594 KiB

Verify the download with SHA-256
21320a87574ca086374bb7b1d3b35f92f2d8ca2ec753d8193b9ea081e24c0f3a

What is included

Task Skill
Create or review system prompts, personas and specialist Agents nuvoca-voice-agent-prompting
Define responsibilities and LLM/Decision routing nuvoca-routing-prompts
Prepare tool descriptions, external API schemas and response processing nuvoca-tool-prompts
Configure knowledge needs and query rewriting nuvoca-query-transformer
Turn sources into thematic Markdown knowledge files nuvoca-knowledge

The package includes shared rules, templates, synthetic examples and local validation scripts. Configuration instructions are written in German. Specify the desired conversation language separately; support depends on the model and voice.

Install in Codex

  1. Download and fully extract the ZIP, including the hidden .agents directory.
  2. Open a terminal in the extracted root directory containing README.md, .agents and plugins.
  3. Register the local marketplace:

    codex plugin marketplace add .
    
  4. Restart Codex. In the plugin directory, install nuvoca-agent-studio under nuvoca, then open a new chat.

  5. Start with an instruction such as:

    Use $nuvoca-voice-agent-prompting. Create an English-speaking reception
    Agent for my company. First clarify its goal, services, boundaries,
    callback workflow and available tools with me.
    

This installation persists across new chats and restarts. Keep the extracted directory in a stable location for the local installation. The command requires a Codex version with plugin support. The official local plugin guide also describes the marketplace structure.

Install in Claude Code

Persistently for your user account

Extract the complete ZIP. Open a terminal in its root directory and copy the entire plugin directory into your personal skills directory:

mkdir -p "$HOME/.claude/skills"
cp -R plugins/nuvoca-agent-studio "$HOME/.claude/skills/"

The plugin then lives at ~/.claude/skills/nuvoca-agent-studio/. Keep its complete contents together, including .claude-plugin/plugin.json, skills/, shared/, contracts/, examples/ and scripts/.

Restart Claude Code, then launch it normally with claude. It loads the plugin on every start; --plugin-dir is no longer needed. Invoke /nuvoca-agent-studio:nuvoca-voice-agent-prompting and describe your task. Organization policy can restrict loading plugins from the skills directory.

To update, replace the complete installed plugin directory with the new version and restart Claude Code. This manually copied installation does not update automatically. See the official persistent installation guide.

For a single session

To try it once, load the extracted plugin directly instead:

claude --plugin-dir /ABSOLUTE/PATH/plugins/nuvoca-agent-studio

Replace the path with your extracted plugin directory. This command loads it only for the session it starts. The included README.md provides further installation instructions.

Give a useful instruction

Provide the business goal, conversation language, target model, existing tools and knowledge bases where known. Attach existing prompts and approved source material. Ask for separate outputs for the relevant Nuvoca fields and realistic test cases.

Drafting does not require MCP. To let the assistant read or save configuration in your workspace, connect Nuvoca separately through MCP. The ZIP contains no account access and does not establish a connection automatically.

Verify the results

Copy reviewed text into the Nuvoca interface or request the change through your MCP connection. Then test normal conversations, follow-up questions and failure cases in a new call.

The optional local validation scripts require Python 3.10 or later. They check format and technical rules; test conversation behavior, retrieval quality and pronunciation with your target configuration. The example prompts help you start.