> ## Documentation Index
> Fetch the complete documentation index at: https://docs.oumi.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# HARNESSES

> Evaluate models in your own agent runtime with Bring Your Own Harness

Use Bring Your Own Harness (BYOH) when you want to evaluate a model inside a harness you already run. Your agent keeps its prompts, tools, memory, and state. You make it available to Oumi at a public HTTPS endpoint, then register that endpoint as a **harness**.

During an evaluation, Oumi sends each assistant turn to your harness and gives it a connection to the model being evaluated. Your harness runs the turn and returns its messages. Oumi can simulate user replies and scores the finished conversations.

![Oumi's evaluation run sends each turn to your harness over the harness protocol, and your harness calls the model under evaluation through the model connection](https://mintcdn.com/oumi/1xlA72C9TkG2ilgC/images/harnesses-connection-light.svg?fit=max\&auto=format\&n=1xlA72C9TkG2ilgC\&q=85\&s=4cb3545a9a1980f3db7f4850187997ae)

Your harness does not need its own access to the model. Each `submit` request includes a model connection with a scoped key, so your agent calls the model under evaluation through Oumi while your prompts, tools, memory, and state stay in your runtime.

![Harnesses page showing registered endpoints and the Register Harness button](https://mintcdn.com/oumi/1xlA72C9TkG2ilgC/images/harnesses-list.png?fit=max\&auto=format\&n=1xlA72C9TkG2ilgC\&q=85\&s=359dfac1278764238902475ac8e204fe)

Once your endpoint implements the [harness protocol](/guides/harnesses/protocol), register it from your project's **Harnesses** page. You can then select it in the evaluation Builder and reuse it across runs.

> [!NOTE]
>
> Harnesses currently support evaluation workflows.

***

## WHAT YOU CAN EVALUATE

| Model | Supported | Notes |
| - | - | - |
| Custom Oumi model | Yes | Oumi serves the pinned model version for the duration of the evaluation |
| API-hosted model (OpenAI, Anthropic, Gemini, Together, Fireworks, OpenRouter) | Yes | Calls go to the provider through Oumi, using your provider key if provided |

Harness evaluations generate new responses for every run. Your harness runs its own tools; Oumi's tool implementations are not used in these evaluations.

***

## NEXT STEPS

> **[Register a harness](/guides/harnesses/register)**
>
> Add your endpoint and credential to a project.

> **[Evaluate with a harness](/guides/harnesses/evaluations)**
>
> Run an evaluation through your harness.

> **[Harness protocol](/guides/harnesses/protocol)**
>
> Implement the routes, payloads, and status codes.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.