LiteLLM

You deploy the proxy into your own infrastructure, self-hosted or air-gapped, and point it at the provider accounts you already have, including internal, fine-tuned and self-hosted models. Applications then call one OpenAI-compatible endpoint, so swapping the backing model is a gateway configuration change rather than a code change in every service. Administrators issue virtual keys scoped by team, project or application, each carrying a budget and RPM/TPM limits; when a key hits its cap, requests stop rather than quietly continuing to spend. Every request is logged and attributed, so spend can be broken down by key, user, team, organisation or tag for chargeback. Routing, load balancing, fallbacks, caching and guardrails are configured at the gateway, and MCP servers and agents are reachable through the same endpoint as the models.

• •One API for 140+ Providers - An OpenAI-compatible interface to 1,800+ models, with day-zero support for new releases, so application code does not change when the model does.
• •Virtual Keys and Hard Budgets - Keys scoped by team, project or app, each with a spend cap and rate limits that actually stop traffic at the ceiling instead of warning after the fact.
• •Spend Attribution - Per key, user, team, organisation and tag, which is what makes internal chargeback possible rather than approximate.
• •Routing, Fallbacks and Caching - Load balancing across deployments, automatic failover when a provider degrades, and response caching, all configured at the gateway.
• •Gateway for Agents and MCP, Not Only Models - Agents and MCP servers are reached through the same endpoint and the same key management as the LLMs.
• •Self-Hosted or Air-Gapped - The gateway runs in your own infrastructure, so prompts, completions and provider keys never transit a vendor's servers.

LiteLLM is for platform and infrastructure teams who have to give a whole engineering organisation model access without losing track of spend or access control, and for anyone who wants their model choice to be reversible. It pays off most where several providers are in play at once, where spend has to be attributed to a team or a cost centre, or where a security review makes a self-hosted gateway the only acceptable shape.

It is a poor fit for a solo developer or a small team calling one provider. A gateway is a service you deploy, monitor, upgrade and page someone about, and the quickstart's one-line install understates that commitment considerably. This is also not a 2026 discovery: it has been running in production at large companies for years, and it earns a place here as an obvious gap rather than as something new. Two caveats belong on the record. The licence is MIT except for the enterprise directory, so the open source claim is partial. And Enterprise publishes no price whatsoever, which is normally what this directory marks a tool down for; only the free self-hosted gateway carries it over the bar.