Choosing models
Pick the model behind the copilot, let emma route automatically, or assign a model to each stage of the pipeline.
The copilot is powered by a large language model, and you can choose which one — or let emma pick for you. This page covers the model picker, automatic routing, and per-stage assignment.
In a nutshell:
- Pick a model for the copilot, or leave it on Auto
- Auto lets emma route to an appropriate model for each request
- Per-stage assignment lets you set different models for different parts of the pipeline
The model picker
Open the model picker in the chat to choose the model behind the copilot. The list shows the chat models your account can access. Your choice applies to the copilot's main reasoning; switch it any time.
Which models appear depends on your plan and any keys your company has added.
Auto
Leave the picker on Auto and emma runs each request on a suitable model — a top-tier model to orchestrate, a lighter one for the delegated and routine work — so you don't have to think about it. Auto is the default and a good starting point.
Per-stage assignment
Under the hood, emma runs a pipeline: an orchestrator that plans, sub-agents that do tool-heavy work, and background tasks like titling and summarizing conversations. You can assign a model to each stage independently — for example, a top-tier model to orchestrate and a lighter one for background work.
If a chosen stage model becomes unavailable (a removed key, say), emma falls back to a sensible default for that stage and tells you.
Which model runs by default, and where your data goes
Two questions every security and AI-governance review asks: which model powers the copilot, and where do your prompts and outputs go.
- By default, the copilot runs on Anthropic's Claude models, through emma's own account. On Auto, the orchestrator runs on a top-tier Claude model and the delegated, tool-running, and background work (titling, summarizing) on a lighter Claude model. The model Auto pins is Anthropic; if you assign models per stage, individual stages can use another provider where that's enabled for your company.
- Every model call is made by emma's servers — emma assembles the prompt and sends it. What your own keys or a self-hosted model change is which model answers, whose credentials are used, and who pays — not whether emma handles the request.
| How you run it | Where inference happens | Credentials | Billing |
|---|---|---|---|
| A default emma model | Anthropic's API | emma's | on your emma plan |
| Your own provider keys | that provider's API | your key, stored encrypted and used only to call that provider | the provider bills you for usage; an emma platform fee may still apply |
| Self-hosted inference | a model server on hardware you run | yours (optional) | you run the hardware |
For data residency: self-hosted inference keeps the model and your data on hardware you control, and your own keys keep usage on your own provider account. In both cases emma still composes and sends the request, so the prompt transits emma on its way to the model. Plan a review around where inference runs and whose credentials are used, not around emma never seeing the request.
Non-Anthropic models and your own keys are available when enabled for your company; the default is Anthropic-only.
What's next
- The copilot — what the model powers
- Bring your own model keys — use your own provider accounts
- Self-hosted inference — run models on your own infrastructure
Last updated on 25 Aug 2026