@nyrra/foundry-ai/docs/usage.md

Thin Palantir Foundry provider adapters and model catalog for the Vercel AI SDK.

# Usage Guide

## What this package is for

`@nyrra/foundry-ai` is for applications that want to use the AI SDK with Palantir Foundry's provider-compatible LLM proxy endpoints instead of calling public provider APIs directly.

Typical use cases:

- local development against a Foundry enrollment
- deployed server workloads that should stay on secure/private Foundry endpoints
- applications that want Foundry governance, rate limiting, attribution, and usage tracking while keeping AI SDK application code

## What is verified today

- env-based server setup with `FOUNDRY_URL` and `FOUNDRY_TOKEN`
- OpenAI, Anthropic, and Google language-model entrypoints from this package
- OpenAI `text-embedding-3-small` and `text-embedding-3-large` embeddings
- application-level routing with AI SDK `createProviderRegistry`

## What is not yet verified

- Palantir TSv1 standalone functions
- Palantir TSv2 standalone functions
- `@osdk/client` or `PlatformClient` fetch integration with this package
- browser/client-side usage

Palantir's docs show the same proxy family used from OSDK and in-platform helpers, but this package should not claim support for those runtimes until it has been validated there.

## Required configuration

| Variable | Required | Purpose |
|---|---|---|
| `FOUNDRY_URL` | yes | Foundry enrollment base URL |
| `FOUNDRY_TOKEN` | yes | bearer token for the proxy endpoints |
| `FOUNDRY_ATTRIBUTION_RID` | no | usage attribution RID header |
| `FOUNDRY_TRACE_PARENT` | no | W3C traceparent value for Foundry observability |
| `FOUNDRY_TRACE_STATE` | no | W3C tracestate value for Foundry observability |

Palantir's [LLM-provider compatible APIs documentation](https://www.palantir.com/docs/foundry/aip/llm-provider-compatible-apis) documents these trace-context headers and notes that third-party OAuth2 applications need the `api:use-language-models-execute` scope.

## Minimal env-based setup

```ts
import { loadFoundryConfig } from '@nyrra/foundry-ai';
import { createFoundryOpenAI } from '@nyrra/foundry-ai/openai';
import { generateText } from 'ai';

const config = loadFoundryConfig();
const openai = createFoundryOpenAI(config);

const result = await generateText({
  model: openai('gpt-5-mini'),
  prompt: 'Summarize why Foundry model aliases are useful.',
});
```

## OpenAI embeddings

Use the same provider instance with AI SDK `embed` or `embedMany`. The friendly aliases are typed convenience constants, and the plain model string is sent as-is to the Foundry embeddings proxy rather than being resolved to a Foundry RID.

```ts
import { loadFoundryConfig } from '@nyrra/foundry-ai';
import { createFoundryOpenAI } from '@nyrra/foundry-ai/openai';
import { embed, embedMany } from 'ai';

const openai = createFoundryOpenAI(loadFoundryConfig());
const model = openai.embeddingModel('text-embedding-3-small');

const { embedding } = await embed({
  model,
  value: 'Foundry-governed embedding input',
});

const { embeddings } = await embedMany({
  model,
  values: ['first input', 'second input'],
});
```

## Local dev and deployed server guidance

- In local development, provide `FOUNDRY_URL` and a valid Foundry token from your normal developer workflow, for example through Developer Console or another approved token source.
- In deployed server runtimes, inject the same values as secrets. Do not expose Foundry tokens to client-side code.
- If your runtime can already surface a Foundry token, base URL, and fetch implementation, treat that as an adaptation opportunity rather than already-verified support.

## Application-level registry composition

The package intentionally does not export a registry helper. Compose one in application code:

```ts
import { loadFoundryConfig } from '@nyrra/foundry-ai';
import { createFoundryAnthropic } from '@nyrra/foundry-ai/anthropic';
import { createFoundryOpenAI } from '@nyrra/foundry-ai/openai';
import { createProviderRegistry } from 'ai';

const config = loadFoundryConfig();

export const registry = createProviderRegistry({
  anthropic: createFoundryAnthropic(config),
  openai: createFoundryOpenAI(config),
});
```

## Repo examples

- [Examples overview with base vs advanced split](https://github.com/shpitdev/foundry-ai/blob/main/examples/README.md)
- [Published base examples](https://github.com/shpitdev/foundry-ai/tree/main/packages/foundry-ai/skills/foundry-ai-provider/references/examples)

## Relevant Palantir docs

- [LLM-provider compatible APIs](https://www.palantir.com/docs/foundry/aip/llm-provider-compatible-apis/)
- [OpenAI Responses proxy reference](https://www.palantir.com/docs/foundry/api/v2/llm-apis/models/openai-responses-proxy)
- [OpenAI Embeddings proxy reference](https://www.palantir.com/docs/foundry/api/v2/llm-apis/models/openai-embeddings-proxy)
- [Anthropic Messages proxy reference](https://www.palantir.com/docs/foundry/api/v2/llm-apis/models/anthropic-messages-proxy)