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

# Deploy LiteLLM Proxy

> Set up LiteLLM Proxy and connect its dependencies with Spring Winter.

## Deploy with an agent

Copy agent prompt

Help me deploy LiteLLM Proxy. Read [https://docs.springwinter.dev/deploy/open-source/litellm](https://docs.springwinter.dev/deploy/open-source/litellm.md), [https://docs.litellm.ai/docs/proxy/deploy](https://docs.litellm.ai/docs/proxy/deploy), and [https://docs.springwinter.dev/api-reference/agents](https://docs.springwinter.dev/api-reference/agents.md). Fetch the public OpenAPI specification from my Spring Winter origin. Use my local SPRINGWINTER\_TOKEN or existing SPRINGWINTER\_API\_TOKEN. Keep credentials out of chat, logs and source.

Confirm my project, AWS connection, Region and budget. Pin a release. Check its required services, container port, health path, network access and durable storage. Data requirements for this deployment: PostgreSQL for virtual keys and database-backed spend tracking. A basic model proxy can run without SQL.

1. If you need virtual keys, create private PostgreSQL first and wait for the writer. Obtain the application connection from Connect.
2. Set DATABASE\_URL to the writer for database-backed features. Set LITELLM\_MASTER\_KEY and provider credentials at runtime. Keep model routing in a versioned config.yaml that references environment variables for secrets.
3. Deploy the pinned proxy image or a GitHub build containing the configuration. Use the configured HTTP port, commonly 4000, and restrict access to administrative routes.
4. Start LiteLLM after its configured database is ready. Let the pinned image initialize its schema. Create scoped virtual keys only after the proxy is healthy.

Show me the planned resources, database and migration order, secret names and estimated cost. Identify components that require separate hosting. Wait for approval before making changes. Use documented Spring Winter operations and wait for each workflow to finish. Send a test completion through one approved model. If virtual keys are enabled, verify that a scoped key can use only its permitted models. Report resource links and anything you could not verify.

## Before you start

[Connect your AWS account](https://docs.springwinter.dev/connect-aws.md) and choose a project. Review the [upstream deployment instructions](https://docs.litellm.ai/docs/proxy/deploy) for the release you plan to pin.

See [upstream virtual key setup](https://docs.litellm.ai/docs/proxy/virtual_keys).

## Data dependencies

PostgreSQL for virtual keys and database-backed spend tracking. A basic model proxy can run without SQL. See [SQL databases](https://docs.springwinter.dev/deploy/databases.md) for private writer access and connection details.

## Deployment order

1. If you need virtual keys, create private PostgreSQL first and wait for the writer. Obtain the application connection from Connect.
2. Set `DATABASE_URL` to the writer for database-backed features. Set `LITELLM_MASTER_KEY` and provider credentials at runtime. Keep model routing in a versioned `config.yaml` that references environment variables for secrets.
3. Deploy the pinned proxy image or a GitHub build containing the configuration. Use the configured HTTP port, commonly 4000, and restrict access to administrative routes.
4. Start LiteLLM after its configured database is ready. Let the pinned image initialize its schema. Create scoped virtual keys only after the proxy is healthy.

Use [Runtime variables](https://docs.springwinter.dev/concepts/environment-variables.md) for runtime credentials. For a [Spring Winter web server](https://docs.springwinter.dev/deploy/web-servers.md), set an app-supported health path that returns HTTP 200 without credentials. Check **History** before running the application check.

## Verify the deployment

Send a test completion through one approved model. If virtual keys are enabled, verify that a scoped key can use only its permitted models.


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