Best LLM gateway: LiteLLM, Portkey, OpenRouter compared

As of 2026-10-14, LiteLLM is the best-documented fit among the three gateways checked here for a self-hosted setup. Portkey's gateway repository documents an MIT license and guardrails, while OpenRouter documents an API endpoint you call directly. The main distinction is a self-hosted proxy versus a service you call by URL.

Quick comparison

How we ranked

Ranking follows documented fit to this query: whether the sources describe the product as a gateway, how much gateway functionality is documented (routing, keys, cost tracking, guardrails), and whether a license is stated. This is an editorial ordering, not a benchmark.

How the options differ

LiteLLM: self-hosted proxy plus Python library

The LiteLLM docs describe an open-source library with a unified interface to call 100+ LLMs using the OpenAI format. The docs also list a self-hosted LLM Gateway (Proxy) with virtual keys, cost tracking and an admin UI.

The proxy is described as a self-hosted OpenAI-compatible gateway, and any client that works with OpenAI works with it without code changes. The docs show it running from a Docker container with a YAML config file.

Gateway features listed include a Router with retry, fallback and load balancing, centralized logging, guardrails and caching. Virtual keys carry per-key, per-team and per-user budgets.

The docs also state that LiteLLM is a unified gateway for LLMs, agents and MCP, with one endpoint for 100+ models, A2A agents and MCP tools.

Portkey: gateway with guardrails and an MIT repository

The Portkey gateway repository describes an AI gateway with integrated guardrails that routes to 1,600+ LLMs with 50+ guardrails through one API. The GitHub API reports an MIT license for the repository.

The Portkey docs show Python, Node and curl usage against a Portkey chat completions endpoint. They also show the OpenAI client pointed at a Portkey gateway URL with Portkey headers.

On deployment, the docs say a feature that does not store request and response bodies can be enabled on request. For enterprises, they say managed hosting can deploy Portkey inside private clouds.

The docs page also carries a banner stating that Portkey is now the Prisma AIRS AI Gateway. The excerpts do not explain what that rename changes.

OpenRouter: one endpoint and SDKs

The OpenRouter quickstart shows requests sent to a chat completions endpoint with a bearer API key. Examples are given in Python with requests, in TypeScript with fetch, and with curl.

The quickstart also lists an @openrouter/sdk package for TypeScript, a Python openrouter client, and an @openrouter/agent package with a tool-calling example. It also lists an MCP endpoint.

The excerpts do not describe OpenRouter's routing behavior, fallbacks, key management or license.

How we compared

We read the primary sources listed above, checked 2026-10-14: the GitHub repository metadata for LiteLLM and Portkey, the LiteLLM getting-started docs, the Portkey introduction page and the OpenRouter quickstart. We did not install, run or benchmark any of the tools, and we did not test latency, reliability or provider coverage.

Limitations

• No pricing is stated in the cited excerpts for any of the three tools. The Portkey page lists a pricing question, but its answer is not in the excerpts.
• The LiteLLM repository metadata reports no recognized license, so reuse terms are not stated in the cited sources.
• Provider counts (100+ and 1,600+) come from each project's own descriptions and are not comparable claims.
• OpenRouter's license, self-hosting options and gateway features beyond the quickstart are not stated in the cited sources.
• Performance claims in repository descriptions were not verified.

Explore the hubs for LiteLLM, Portkey, OpenRouter and the Agent infra category. Newest same-tool launch: LiteLLM v1.105.0 adds MCP protocol version settings. Moving off LiteLLM? Read LiteLLM alternatives open source: Portkey and OpenRouter.