qtap-plugin-mistral-ai

Mistral AI support for Quilltap

Plugin 1: Mistral AI (qtap-plugin-mistral)

Summary

Connect to Mistral's API at api.mistral.ai/v1 using the @mistralai/mistralai SDK or the OpenAI SDK with base URL override. Mistral offers the strongest value proposition of any new plugin: reasoning models, vision, tool calling, OCR, audio transcription, and the cheapest high-quality models in the market.

API Details

| Field | Value | | -------------- | ------------------------------------------------------------- | | Base URL | https://api.mistral.ai/v1 | | Auth | Authorization: Bearer {MISTRAL_API_KEY} | | SDK | @mistralai/mistralai (preferred) or openai with baseURL | | API format | Chat Completions (/v1/chat/completions) | | Streaming | SSE via stream: true | | Docs | https://docs.mistral.ai/api |

Supported Capabilities

| Capability | Support | Notes | | ------------------------- | ------- | ------------------------------------------------------------------------------------------------------ | | Chat completions | ✅ | Standard chat completions format | | Streaming | ✅ | SSE with stream: true | | Tool/function calling | ✅ | Standard OpenAI format; tool_choice: "any" forces tool use; parallel_tool_calls supported | | Vision/image input | ✅ | image_url content parts (URL or base64); supported on Mistral Small 3.1+, Pixtral, Mistral Medium 3+ | | JSON mode | ✅ | response_format: { type: "json_object" } and { type: "json_schema" } (structured outputs) | | Reasoning models | ✅ | Magistral models with reasoning_effort param ("high", "medium", "none") | | Embeddings | ✅ | mistral-embed and codestral-embed models | | Image generation | ❌ | Not available via API | | Web search | ⚠️ | Available via Agents API connector (server-side), not standard chat completions | | File attachments | ✅ | Via vision (images) and Document Library (PDFs, for Agents API) |

Models (current as of March 2026)

Chat/General:

• mistral-large-latest (Mistral Large 3) — flagship, 675B MoE, 128K context
• mistral-medium-latest (Mistral Medium 3.1) — GPT-4 class, $0.40/$2.00 per M tokens
• mistral-small-latest (Mistral Small 4) — hybrid instruct/reasoning/coding, 256K context, multimodal
• ministral-3b-latest, ministral-8b-latest, ministral-14b-latest — edge/small models

Reasoning:

• magistral-medium-latest — reasoning with configurable effort
• magistral-small-latest — lighter reasoning model

Code:

• codestral-latest — code-optimized
• devstral-latest (Devstral 2) — agentic coding

Specialized:

• pixtral-12b-2409 — vision-first model
• mistral-ocr-latest — document OCR (not chat completions; separate endpoint)
• voxtral-small-latest — audio chat
• voxtral-mini-latest — audio transcription

Plugin-Specific Implementation Notes

1. SDK choice: Use @mistralai/mistralai for the primary implementation. The Mistral SDK handles token counting, streaming, and tool calling natively. The OpenAI SDK works as a fallback but misses Mistral-specific features like reasoning_effort and safe_prompt.

2. Reasoning model detection: Check for magistral prefix in model name. When detected:

   • Support reasoning_effort parameter (map from Quilltap's UI toggle)
   • Stream reasoning content separately if the SDK exposes it

3. Vision support: Use the standard image_url content type. Detection: models with pixtral, mistral-small (3.1+), mistral-medium (3+), or mistral-large (3+) in the name support vision.

4. safe_prompt parameter: Mistral supports a safe_prompt: true option that injects a safety preamble. Expose this in the connection profile as a toggle (default: false for roleplay use cases).

5. prefix feature: Mistral supports assistant message prefilling via a prefix parameter. This is useful for locking character voice. Expose in connection profile advanced settings.

6. Cheap LLM candidates: ministral-3b-latest, ministral-8b-latest, mistral-small-latest

7. Embedding models: Register mistral-embed and codestral-embed as embedding providers.

Estimated Scope

• ~400 lines for provider.ts (based on existing OpenAI plugin complexity)
• ~50 lines for index.ts
• Mistral SDK is an additional dependency (~2MB bundled)