litellm - 💡(How to fix) Fix Feature Request: support per-model SOCKS5 / outbound proxy settings in LiteLLM Proxy [1 participants]

What problem is this solving?

Today LiteLLM Proxy appears to support outbound proxying mainly via process-level / environment-level proxy settings such as:

• HTTP_PROXY
• HTTPS_PROXY
• ALL_PROXY
• AIOHTTP_TRUST_ENV

That works for a single global egress path, but it does not solve a common deployment need:

Different upstream models/providers need to use different outbound proxies.

Examples:

• some model deployments must go through a SOCKS5 proxy
• some should use direct outbound traffic
• some should use a different HTTP/SOCKS proxy depending on provider, region, compliance, or network topology

Right now the practical workaround seems to be running multiple LiteLLM Proxy instances and splitting traffic externally, which adds operational complexity.

Requested feature

Please support per-model outbound proxy configuration in config.yaml for LiteLLM Proxy.

For example, something like this:

model_list:
  - model_name: openai-via-socks5
    litellm_params:
      model: openai/gpt-4o
      api_key: os.environ/OPENAI_API_KEY
      outbound_proxy: socks5://127.0.0.1:1080

  - model_name: anthropic-direct
    litellm_params:
      model: anthropic/claude-sonnet-4-20250514
      api_key: os.environ/ANTHROPIC_API_KEY
      outbound_proxy: null

  - model_name: azure-via-http-proxy
    litellm_params:
      model: azure/gpt-4o
      api_base: os.environ/AZURE_API_BASE
      api_key: os.environ/AZURE_API_KEY
      outbound_proxy: http://proxy.internal:8080

Even if the final config name is different (proxy, transport_proxy, network_proxy, etc.), the key need is:

• configurable per deployment / per model entry
• support for SOCKS5 and regular HTTP(S) proxies
• ability to explicitly override global env-based proxy behavior

Why this matters

This would make LiteLLM Proxy much easier to use in real-world environments where outbound routing differs by model/provider:

• geo/network routing by provider
• regulated environments
• proxy segmentation between tenants/providers
• local lab / bastion / tunnel setups
• migration scenarios where only some providers are reachable through a proxy

Expected behavior

Ideal behavior

1. If a per-model outbound proxy is set, requests for that model use it.
2. If a per-model outbound proxy is explicitly null / disabled, the request bypasses proxy env vars where possible.
3. If not set, current global env-based behavior remains unchanged.

Nice to have

• support in both sync + async paths
• works whether LiteLLM is using aiohttp transport or httpx
• docs include examples for SOCKS5
• validation error if the proxy scheme is unsupported

Workaround today

The current workaround is to run multiple LiteLLM proxy instances, each with different env proxy settings, and route models across instances manually. It works, but it is much more operationally expensive than a native per-model setting.

If helpful

I’d be happy with even a first version that:

• supports per-model proxy config for the proxy server path only
• supports SOCKS5 via ALL_PROXY-equivalent client configuration internally
• leaves advanced routing/policy to future iterations

Thanks — this would remove a lot of deployment pain.