Models and connectors

Every model you use, behind one set of contracts.

Cloud APIs, subscriptions and engines that run inside the app all speak the same contracts. Pick a connector per session, and change it whenever you like.

Chat and agents

Text models.

Images attach to prompts on every connector that accepts them.

OpenAI

API key, through the Responses API.

Anthropic

API key, through the Messages API.

Google AI Studio

Gemini API key.

Google Vertex AI

Gemini and Claude models in your Google Cloud project.

Amazon Bedrock

Converse through a named AWS profile.

Bedrock AgentCore

An AWS-hosted AgentCore harness session.

OpenAI-compatible

Any endpoint that serves Chat Completions: OpenRouter, Groq, DeepSeek, xAI, Kimi, MiniMax, Ollama, llama.cpp and more.

Apple on-device

Apple Foundation Models on the Mac itself.

Self-hosted Qwen

GGUF Qwen models run inside the app on the Tensor engine.

Subscriptions

Use the plans you already pay for.

Darkspyre runs on your existing Claude, ChatGPT and Gemini subscriptions as full agent sessions, held to the session's permission level.

  • Claude. Through the official Claude CLI, signed in separately. Darkspyre never copies its login.
  • ChatGPT. Through a Codex sign-in that Darkspyre manages.
  • Gemini. Through the Antigravity CLI, signed in separately.
The Add connection dialog listing OpenAI, Anthropic, ChatGPT, Claude and Gemini subscriptions, Apple on-device, Bedrock, Google AI Studio, Vertex AI, AgentCore and the built-in Tensor engines
Settings, Providers, Add connection.

Beyond text

Images, speech and audio.

KindConnectorRuns
Image generationComfyUI workflowsYour ComfyUI server
Image generationSupraInside the app
Speech synthesisPocketTTSYour PocketTTS server
Audio generationTangoFluxYour TangoFlux server
Speech transcriptionWhisper (base.en to large-v3-turbo)Inside the app

Utility models

Embeddings and decisions.

Embeddings

OpenAI and any OpenAI-compatible embedding endpoint.

Decisions

Laya answers typed yes-or-no questions on ONNX Runtime inside the app. RightSize uses it to size work before it is dispatched.

Local engines

Local models with no Python.

TensorSharp and ONNX Runtime are built into the app. Qwen text, Supra images, Whisper transcription and Laya decisions run in process, with nothing else to install.

  • Same contracts. A local engine answers through the same interfaces as the cloud connectors, so tools and sessions do not care where a model runs.
  • Discovered by kind. The app finds installed engines for each model kind and offers them in Settings.

Switching

Change providers, keep the conversation.

Move a session from one model or provider to another at any point. Text, tool calls, tool results and attachments carry over. State that only one provider understands, such as signed reasoning, stays behind and is counted.

Agent sessions carry over too. Codex and Bedrock take the history as structured context. Claude, AgentCore and Antigravity receive a one-time transcript of the conversation so far.