Shisa AI Model Provider for Dify

A Dify model-provider plugin for connecting Dify applications directly to Shisa AI's Japanese and English APIs.

This repository is maintained by Shisa AI. It does not imply certification, endorsement, or support by Dify unless separately stated by Dify.

Supported interfaces

Translation is exposed separately through Shisa AI Tools for Dify, because Dify does not provide a standard Translation model-provider interface.

Requirements

• A Shisa AI account and API key from Shisa Platform
• A Dify installation that supports model-provider plugins
• Python 3.12 for local development

Current prices, quotas, model availability, and account-specific rates can change. Check Shisa Platform and the official API documentation for current information rather than relying on repository snapshots.

Installation

1. Download from the v1.0.3 GitHub Release.
2. Optionally verify its GitHub-provided SHA-256 digest and provenance attestation.
3. In Dify, open Plugins, choose installation from a local package, and upload the file.
4. Open the Shisa AI model-provider settings and enter your Shisa AI API key.
5. Select the required Shisa LLM, ASR, or TTS model in your Dify application.

Only the file is installable in Dify. GitHub’s automatically generated Source code (zip) and Source code (tar.gz) archives are not plugin packages.

Do not commit API keys or Dify remote-debug credentials.

Dify behavior and limitations

• Standard Dify TTS uses MP3 because Dify's application audio endpoint expects .
• Dify Workflow TTS and Tool nodes buffer completed files before downstream nodes receive them; they do not provide progressive workflow playback.
• Some Dify surfaces can retrieve voices dynamically, while workflow selectors may rely on packaged fallback metadata.
• Actual TTS formats and streaming support vary by voice. is the source of truth.
• Silent or wrong-microphone input may be returned by the ASR service as the exact marker ; this plugin converts only that exact case-insensitive marker to an empty transcript.

For native audio formats and dynamic voice tooling, use the separate Shisa AI Tools plugin.

Workspace-wide ASR defaults

The provider settings expose the documented optional Shisa ASR parameters: language, hotwords, temperature, , frequency penalty, repetition penalty, and VAD. Dify does not pass these values to the standard Speech-to-Text interface per request. Instead, the plugin reads explicitly configured values from the provider credentials and adds them to each ASR request.

Scope warning: these settings are workspace-wide provider defaults. They affect every Dify application in the workspace that uses this Shisa ASR provider. They are not scoped to one Chatflow. Leave a field blank to omit it and use the Shisa API default.

Hotwords accept either a JSON string array:

or a comma-separated list. For per-node settings, use Transcribe Audio in the separate Shisa AI Tools plugin.

Development

On Windows PowerShell:

Use the debug host and temporary key shown in the Dify plugin-debug dialog. Never publish , , application tokens, or remote-debug keys.

Package

With the Dify plugin CLI installed:

Protected tags invoke . CI builds and validates the , generates a runtime SBOM, attests provenance, and publishes only the installable package to the matching GitHub Release. Generated release files are intentionally excluded from Git.

Service endpoints

• LLM:
• ASR:
• TTS:
• Voices:

License

Licensed under the Apache License 2.0.

Security and privacy

• Report vulnerabilities according to SECURITY.md.
• Data-handling details are documented in PRIVACY.md.
• Contributions are described in CONTRIBUTING.md.