narova-mimo

Narova + Xiaomi MiMo (V2.5-TTS)

Use this skill as an optional external speech provider for Narova. Keep every Xiaomi MiMo API detail in this companion and communicate through the generic narova-tts-provider/v1 JSONL protocol. Let Narova own staging, provenance, generation recipes, the asset registry, caching, and rendering.

Requires the standalone Narova CLI, the Narova skill, Python 3.10+, network access, a Xiaomi MiMo platform account with API access, and MIMO_API_KEY.

Setup

1. Install the standalone Narova CLI as described by the narova skill, then install narova and narova-mimo as independently selected skills from ammar-hasan/narova.
2. Locate this installed skill from the directory containing this SKILL.md; call it <narova-mimo-skill-dir>.
3. Verify the worker:

bash <narova-mimo-skill-dir>/tool/setup.sh

1. Set the key only in the environment that runs Narova:

   export MIMO_API_KEY="..."
   

2. Explicitly register the provider and verify:

   narova providers add \
     <narova-mimo-skill-dir>/tool/provider.json
   narova providers doctor mimo
   

Installing this skill registers no provider by itself. Unregistering the provider does not remove the skill.

Read references/configuration.md before use. It documents endpoint selection, the three model identities (preset voices, voice design, voice cloning), option mapping, the voice-design anchor pattern, errors, security, billing, AI-generation disclosure, and removal.

Operating rules

• Treat synthesis as an explicit, billed network operation. The authored text is sent to Xiaomi MiMo.
• Keep MIMO_API_KEY only in the process environment or a user-managed secret store; never put it in reel.config.mjs.
• Do not add a Xiaomi SDK to Narova or modify Narova's setup. This worker uses the Python standard library.
• Do not copy this worker into the Narova skill. Registration composes two independently installed skills.
• Do not automatically retry a failed request. A lost response may still have been processed and billed.
• Let Narova publish synthesized audio only after worker success; worker failures leave assets and the registry unchanged.
• Disclose that speech output is AI-generated where that matters to the audience. Cloning a real person's voice is the author's consent and rights responsibility.