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

# Use Whitney with a coding agent

> Give an agent the API reference, cookbooks, and Whitney implementation skill.

The docs site exposes its context through Mintlify's public MCP endpoint and
installable skill discovery.

## Prompt an MCP-enabled agent

Add this docs site as an MCP server using:

```text theme={null}
https://staging.docs.trywhitney.com/mcp
```

Then prompt your agent with a concrete algorithm and constraints:

```text theme={null}
I want to run DAPO on Whitney. Read the DAPO section in the algorithms
cookbook and examples/training/http/dapo.py — it already has a real
build_loss. Adapt its reward() function for <describe your task>, then run
it with rl_loop.py. Use provider=modal or tinker (not auto, so the
capability check is deterministic), keep durable operation polling and
stable idempotency values on retry, and always cancel/close on failure. Do
not import a provider SDK or submit a managed recipe.
```

## Install the Whitney skill

Agents that understand the open skills format can discover the skill from the
docs domain:

```bash theme={null}
npx skills add https://staging.docs.trywhitney.com
```

The skill directs the agent to capabilities, API schemas, primitive ordering,
the requested algorithm cookbook, and safe cleanup behavior. You still provide
`WHITNEY_BASE_URL` and `WHITNEY_API_KEY` through your local secret manager;
never paste a real key into an agent prompt or repository.

## What to ask for

State the algorithm, model constraints, data or reward callbacks, desired
checkpoint or artifact behavior, and maximum cost if you need an exact ceiling.
When you do not care about provider routing, ask for `provider: auto`.

Implement algorithms as loops over Whitney HTTP primitives. There is no Whitney
SDK requirement for v1 and no managed recipe endpoint.