Bifrost is an open-source, high-performance AI gateway from the team at Maxim. It gives applications a single OpenAI-compatible API in front of more than 10,000 models, and doubles as an LLM gateway, MCP gateway and agent gateway. The project has gathered about 9,000 stars on GitHub and is under very active development in 2026.
Written for speed, Bifrost adds under 200 microseconds of overhead even at 5,000 requests per second, according to the project. In front of providers such as OpenAI, Anthropic and many others, it provides automatic failover and load balancing so traffic survives a provider outage, semantic caching to cut repeat inference costs, and governance features including virtual keys, budgets, guardrails and observability for LLM spend. It works with any existing SDK or agent framework through its OpenAI-compatible endpoints, and its MCP gateway mode — including MCP Code Mode — lets agents reach tools through the same layer. Bifrost can be deployed as a standalone binary, in Docker or on Kubernetes, and is configured through an API and web UI, which makes it practical to roll out as shared infrastructure for a whole engineering organization. Because it speaks the OpenAI format in both directions, existing applications usually switch to it by changing a base URL — no SDK rewrite, and no lock-in if you later move providers.
Bifrost’s core gateway is free and open source under the Apache-2.0 license for self-hosting. Maxim also offers paid enterprise options with advanced governance and support for organizations that need them.
Bifrost is for engineering teams running LLM features in production who want one fast, governable gateway instead of separate integrations — and separate failure modes — for every model provider.