Give your AI Agent local voice tools, one verified job at a time

Choose a task you can inspect

Start with “read this passage and save the audio.” Once that works, ask for a course introduction, voices for a dialogue, or a searchable transcript of a recording.

A first request that also changes voices, writes a long script and imports external video mixes connection, model and file-access problems. Keep the first task small enough to identify which step succeeded.

Connect in the app, test in the Agent

1. Open Sonano, check that its engine is running, and enter Connect Agent.
2. Choose your Agent and follow the current actions for the Skill and MCP setup. Update an older configuration when offered.
3. Enable tools, restart the host or open a new session as required.
4. Copy the app’s test message into the selected Agent’s conversation.

A saved configuration is preparation. Verify that the Agent can discover Sonano, inspect ready models and finish a small job. Host settings and configuration locations differ; follow the app’s host-specific guidance.

Discover before generating

This illustrative request explains the sequence. It has not been executed here; for a first connection, prefer the app’s generated test message.

Check Sonano’s service status and list ready speech models and suitable English narration voices. If a compatible combination is available, generate “Check the voice before starting the production” as a short audio file, then report the job status and artifact location. Report missing models or voices first.

Use discovered model and voice identifiers rather than guessing them from an article. Follow the job to a clear completed or failed state. A generated artifact that can actually be accessed is the result you need.

Tool access and file access are different

MCP exposes discoverable audio tools; the Skill explains workflows. Sonano’s current MCP connection uses a host-launched stdio process. Its ordinary HTTP engine address is not an HTTP MCP address.

External media still depends on host permissions and the available transport. Ask which file the Agent can access, and follow the current Skill’s upload guidance. Model support for transcription does not grant access to arbitrary disk paths. Audiobooks have app and HTTP workflows, but are not automatically available as MCP tools.

Check the job, file and sound

Inspect jobs or history in Sonano and use Request Audit to review calls. Play the delivered audio and check its words and voice. Visible history also depends on your external-call recording settings, so an empty history alone does not prove no request occurred.

If the Agent uses a cloud language model, text may still pass through that provider. Local speech generation describes the audio inference step. Continue with the Agent page and connection documentation.