🌬️ hivekeep-plugin-mistral

Mistral AI provider for Hivekeep. Brings Mistral's chat models (mistral-large-*, mistral-medium-*, mistral-small-*, pixtral-* for vision, …) and Voxtral speech-to-text (voxtral-mini-2507, voxtral-small-2507) to Hivekeep with native tool calling, streaming, and vision (Pixtral family).

Install

In Hivekeep's web UI:

1. Settings → Plugins → Browse → npm tab
2. Search for mistral
3. Click Install on hivekeep-plugin-mistral
4. Settings → Providers → Add provider → Mistral AI
5. Paste your API key from console.mistral.ai/api-keys

Alternative: install from git instead of npm via Settings → Plugins → Install from git with https://github.com/MarlBurroW/hivekeep-plugin-mistral as the URL.

What this plugin does

Contributes two native providers to Hivekeep's provider registry, both under type: 'mistral' so a single API-key row covers chat + transcription (same pattern the host's built-in OpenAI provider uses across LLM / Embedding / Image / TTS / STT):

• LLMProvider — chat against /v1/chat/completions
• STTProvider — Voxtral against /v1/audio/transcriptions

From there the model picker on every Agent sees Mistral models alongside Anthropic / OpenAI / Gemini built-ins — no special-casing.

Chat capabilities

• ✅ Streaming chat via SSE (POST /v1/chat/completions)
• ✅ Native function / tool calling (tools + tool_calls)
• ✅ Vision (Pixtral and recent large/medium models — images passed as data: URLs)
• ✅ Live model catalogue (GET /v1/models) — embedding / moderation / OCR models filtered out
• ✅ Schema-driven capability detection — when Mistral adds a new model with capabilities.function_calling: true, it just works

Voxtral STT capabilities

• ✅ Two production models: voxtral-mini-2507, voxtral-small-2507
• ✅ Language hint (language form field, ISO 639-1)
• ✅ Auto-detected language returned in the response
• ✅ Segment-level timestamps when the host asks for them (timestamp_granularities[]=segment)
• ❌ Diarization (speaker labels) — Voxtral doesn't ship it; diarize hints arrive as a warning
• ❌ Prompt biasing (Whisper-style vocabulary biasing) — no prompt field on the upstream API

Out of scope (for now)

• Embeddings (mistral-embed) — could be a third provider in this plugin; raise an issue if you need it
• Moderation (mistral-moderation-latest) — Hivekeep doesn't have a moderation primitive yet
• OCR (mistral-ocr-*) — not yet a Hivekeep primitive
• Codestral fill-in-the-middle — completion-style, not chat

Development

git clone https://github.com/MarlBurroW/hivekeep-plugin-mistral
cd hivekeep-plugin-mistral
bun install
bun test

To iterate against a local Hivekeep install: drop this directory (or a symlink to it) into Hivekeep's plugins/ folder and reload the plugins from the UI. The plugin loader hot-reloads on file change.

Permissions

The plugin declares a single permission in its manifest:

• http:api.mistral.ai — required for chat / model-listing / transcription requests

No vault, storage, cards, cron, or agents permissions needed.

License

MIT