Models and Providers

Models and Providers

Eva separates the user experience from the model route. Choose the route based on where you want inference to happen, what credentials you have, and whether you need Copilot-backed tools or local-only operation.

At a Glance

Route	Best for	What you need
Eva (AIG)	Memory, retrieval, persona, goals, skills, and optional review	A backend: Copilot ACP, OpenAI, GitHub Models, or LM Studio
Copilot ACP	Copilot CLI models, MCP retrieval, subagents, coding workspaces	Copilot CLI and a local copilot auth login
OpenAI	Direct OpenAI chat, images, TTS, embeddings	OpenAI API key
GitHub Models	Direct models through GitHub Models	GitHub PAT with Models permission
Google Gemini	Gemini compatibility route	Google Gemini API key
LM Studio	Local OpenAI-compatible inference and local-only operation	A locally running LM Studio model

Recommended Default: Eva (AIG)

Choose Eva (AIG) in the model menu for the integrated route. AIG can:

• assemble relevant local SQLite or optional Kusto memory;
• retrieve current data when the request needs it;
• apply the Eva persona and your configured system prompt;
• route the final model call to the selected backend;
• reflect after a response to update relevant memory and state; and
• optionally request a focused independent review for consequential work.

The model is still important. AIG gives Eva a consistent experience around the chosen backend; it does not make every provider have identical capability or pricing.

Provider Setup

Copilot ACP

Install and authenticate GitHub Copilot CLI on the host:

copilot auth login

Select a Copilot ACP model in Settings. This route powers Copilot-backed MCP retrieval, isolated subagents, and the initial Coding Workspaces runtime. The available model catalog evolves with the Copilot CLI and your entitlement.

Direct OpenAI

Add an OpenAI API key in Settings > Auth, then choose an OpenAI model or an OpenAI-backed AIG backend. Direct OpenAI preserves Eva's memory, persona, response rendering, action markers, and adaptive review pipeline. It also enables OpenAI image generation, TTS, transcription, and semantic recall where configured.

GitHub Models

Add a GitHub PAT with Models permission in Settings > Auth. For private workspace imports, the token also needs repository read access to the selected repository. Use the least privilege needed and revoke the token when you no longer use the route.

Google Gemini

Add a Gemini API key in Settings > Auth and select the compatible Gemini model. The route follows the same session and post-response reflection lifecycle as the other direct-provider routes after a successful answer.

LM Studio

Start a local model server in LM Studio, choose an LM Studio model, and use Local data mode. Eva talks to its local OpenAI-compatible endpoint. No cloud model credential is required for this inference route. See Local Mode and MCP for local retrieval and tool setup.

Model Controls

Settings > Models provides controls such as temperature, max output tokens, reasoning effort, the AIG backend, ACP model, and adaptive review. Not every provider accepts every parameter. Eva applies known provider-specific rules, including omitting unsupported controls for some reasoning models.

Higher reasoning effort can improve difficult work, but it can also increase latency and premium usage. Start with the model default and increase only when a task benefits from more deliberate reasoning.

Cloud and Local Data Modes

Mode	Retrieval path	Typical use
Cloud	Copilot ACP and configured MCP servers	Copilot-backed current-data research and tools
Local	LM Studio plus Eva-managed local MCP subprocesses	Local inference and direct local tools

Eva restores the most recently selected mode on startup. Local mode does not make every integration offline: browser automation, configured cloud memory, or an external service still behaves according to its own configuration.

Cost and Privacy Decisions

1. Use LM Studio when inference should remain local.
2. Use direct provider APIs when you want that provider's features and are comfortable with its service terms.
3. Use Copilot ACP when you need Copilot CLI-backed tools, subagents, or coding workspaces.
4. Configure one route first, verify it, then add optional services.

Read Memory, Learning, and Privacy before deciding which data should be retained locally or sent to configured services.