What Bifrost is
Bifrost is an AI gateway written in Go by Maxim AI (legal entity H3 Labs Inc.). The core is Apache-2.0 with no carve-outs (LICENSE). You run it yourself: a single binary started with npx -y @maximhq/bifrost, a Docker image, a Helm chart, or embedded in your application as a Go package. There is no hosted Bifrost. The paid Enterprise tier is a licensed feature set plus support for your own deployment (VPC, on-premise, or air-gapped), with quote-based pricing and no public dollar amounts.
It launched mid-2025 and moves fast. As of August 16, 2026 the repo has roughly 7,300 stars, the latest release (transports v1.6.11) is two days old, and the docs list about 21 providers, with the README claiming 23+ providers and 1,000+ models. It exposes native /openai and /anthropic endpoints plus a Google GenAI surface, with documented drop-in replacement for those SDKs, LangChain, and the LiteLLM SDK.
Where Bifrost is the right call
The free core is unusually complete. Automatic failover, weighted load balancing, virtual keys with budgets and rate limits, custom routing rules, semantic caching, an MCP gateway with tool governance, and a built-in observability UI all ship under Apache-2.0 — several of these are paid features in competing gateways.
Performance is the headline pitch. Their self-published benchmark reports 11 µs internal overhead at 5,000 RPS and a 54x p99 advantage over LiteLLM at 500 RPS. Treat the numbers with care: they are undated, the LiteLLM comparison ran on 2-vCPU instances where Python gateways degrade worst, the 11 µs figure comes from a larger 4-vCPU t3.xlarge (the 2-vCPU box measured 59 µs), and the headline multiplier drifts between 40x, 50x, and 54x across their own pages, with no independent replication found. The architectural point survives the caveats — a compiled Go binary avoids the failure class those tests target, and they publish a benchmark page and a run-your-own guide.
Self-hosting is also the strongest data boundary available. Bifrost inside your perimeter, with content logging turned off, puts no third party in the request path. No hosted gateway can match that, ours included.
Three differences that decide it
Who runs the gateway, and how you verify it
With Bifrost, assurance is your own audit plus your own operations: read the Apache-2.0 source, deploy it, keep it patched and available. Clustering and high availability are Enterprise, so HA is quote-priced or engineered yourself. TrustedRouter is hosted, and you cannot inspect our servers, so we attest instead: the gateway runs published builds inside TEEs on GCP Confidential Space, AWS Nitro Enclaves, and Azure Confidential Containers, each with a live attestation endpoint (api, api-aws, api-azure.trustedrouter.com) and a verifier script — details at trust.trustedrouter.com and /security. The boundary is precise: attestation covers our gateway build and nothing downstream. Model providers are covered by cited retention-policy tiers, except on trustedrouter/e2e routes (tinfoil, phala), where the provider's own confidential-compute and end-to-end encryption posture applies — the provider's mechanism, not our attestation.
What happens to prompt content by default
Bifrost's gateway mode enables its logging plugin by default and writes full request and response content to a local SQLite store. A disable_content_logging flag exists, and it defaults to false — the behavior is documented in their observability docs. It stays on your own disk rather than with a vendor, but teams expecting a pass-through proxy should flip the flag first. On TrustedRouter, realtime inference keeps no durable prompt or output logs; we retain operational metadata only (request ids, model, token counts, latency, cost, region, API-key hash) — see /privacy. The opt-in Batch API is the exception: enclave-encrypted artifacts retained up to 30 days.
Provider keys, catalog, and what routing costs
The fee bases differ, so keep them apart. Bifrost software costs nothing to license; you hold a key for each provider, pay each provider directly at their price, and carry the infrastructure and operations cost, with Enterprise features quote-priced. TrustedRouter is one prepaid key across 550+ model routes and roughly 49 providers (as of August 2026), billed at the provider's token price plus 5.5% with a $0.01 per million token floor; video is the provider quote plus 20%. Per-route prices are on /models, the full schedule on /pricing. BYOK is supported; we do not publish a separate BYOK fee. Routing also carries privacy tiers Bifrost does not model: trustedrouter/zdr restricts to providers with contractual zero data retention, trustedrouter/e2e to confidential-compute providers with provider-side encryption, and trustedrouter/eu to an EU provider order, all composable with per-request preferences.
What we do not claim
- Our gateway and control plane are not open source. They are source-available under BUSL-1.1, converting to Apache-2.0 four years after each release; the SDKs are Apache-2.0 or MIT. On licensing, Bifrost's core is more open than ours today.
- Our attestation does not extend to model providers, and community user-* routes leave the attested boundary entirely.
- We are younger than they are. Our repos have been public since late April 2026 and our benchmark history starts June 2026. Judge our record by the status page and the frozen monthly benchmark reports, not by our word.
- We publish a DPA, BAA, and subprocessor list but claim no SOC 2 or HIPAA certification today. Maxim AI displays SOC 2 and ISO 27001 badges for its company and hosted platform; a self-hosted Bifrost deployment's compliance is the operator's own in either case.
Migration reality
For OpenAI-format clients the move is a base URL and API key change — Bifrost's own drop-in story running in reverse. Clients on Bifrost's /anthropic endpoint can target our Anthropic-style Messages surface; Google GenAI-native clients need code changes, because we expose OpenAI and Anthropic formats. Provider keys re-home: retire them on prepaid, or keep committed-spend discounts through BYOK. Bifrost virtual keys, per-team budgets and rate limits, routing rules, semantic-cache settings, MCP tool configs, and plugin hooks have no automatic mapping — recreate what has an equivalent (MCP: /docs/mcp); plugins do not carry over. Export Prometheus and SQLite/Postgres log history before switching; our observability is metadata analytics built from traffic routed through us. You shed gateway operations: upgrades, scaling, and the HA engineering Bifrost prices into Enterprise.