Open-source AI gateway you host yourself
INFRO vs Bifrost
Bifrost is Maxim AI's open-source AI gateway, Apache-2.0 licensed and built to be self-hosted. Its open-source edition carries more than most gateways put in a paid tier: virtual keys with dollar budgets and rate limits at key, team and customer level, fallbacks and weighted load balancing across 20+ providers, semantic caching, an MCP gateway with per-key tool allow-lists, OpenTelemetry, and pricing overrides for your own rates. Guardrails, SSO, role-based access, audit logs, clustering and in-VPC deployment come with its Enterprise edition. If you want a capable gateway you run yourself at no licence cost, it is a serious answer.
- You want to run the gateway in your own infrastructure, from open-source code, with no vendor agreement. That is exactly what Bifrost is for; INFRO's self-hosted and hybrid deployments come only with an enterprise agreement, and its gateway is not open source.
- You want to start at zero licence cost. Bifrost's open-source edition includes budgets, virtual keys, semantic caching and an MCP gateway; INFRO's entry plan is paid, with a trial.
- Semantic caching matters to your traffic. Bifrost serves similar prompts from its cache; INFRO's response cache answers exact repeats of chat completions only.
- You want to read, patch or extend the code — Go or WASM plugins in the request path — rather than configure a service.
- You need it working today. It is available now; INFRO is in private early access.
This section is not a courtesy. If one of these describes you, pick Bifrost — a customer who adopts the wrong layer on our advice is worse for us than one we never had.
Where the two actually diverge.
- Software you run, or a service you use
- Bifrost is yours to operate: deployment, upgrades, scaling, and the on-call that comes with a component every model request passes through. Its Enterprise edition adds clustering and in-VPC deployment; neither its documentation nor its pricing page describes a hosted service. For a platform team that already runs this kind of infrastructure that is the right trade. INFRO's plans are hosted, so there is nothing to deploy, and the mirror image holds: your requests pass through our service rather than yours unless an enterprise agreement puts INFRO in your own infrastructure.
- The estimate, and then the provider's own number
- Both products estimate cost from a price catalog and both take your own rates — Bifrost through pricing overrides scoped from global down to a single virtual key and provider key, INFRO through rates stored per model and provider kind from the Team plan. The difference is what happens after. On the Business plan INFRO reads OpenAI's and Anthropic's own cost APIs with a separate admin credential, stores what the provider reported beside what INFRO estimated, and states the variance. Bifrost's documentation describes no equivalent. And where no price is known, the two differ in default: Bifrost's docs note that a custom provider with no catalog price costs $0 until you add an override, while INFRO records such a request as unpriced, and a blocking budget refuses unpriced traffic unless you allow it.
- Beyond the gateway
- INFRO is a gateway plus the operating surface around it: a console where finance and team leads read spend by project, key and end user; a weekly summary and monthly report emailed to owners and finance members; spend anomaly and drift findings by email, and from the Startup plan in Slack, Teams, PagerDuty or a signed webhook; image, video and audio behind one request shape each, with long renders as signed-webhook jobs; and INFRO Studio, a workspace for image, video and audio work on your own provider connections, on every plan. None of that is a gateway feature, and a team that only needs a gateway should not pay for it.
- Not either-or
- A Bifrost deployment you already run speaks the OpenAI API, so it can sit behind INFRO as an OpenAI-compatible connection: keep it for the providers and plugins only it reaches, and put INFRO's attribution, budgets and reconciliation in front of the whole estate.
Side by side.
Our column carries dashes too. A comparison where one side is all checks is one nobody believes, and rightly.
| Capability | Bifrost | INFRO |
|---|---|---|
| One API across many providers | yes20+ providers | yesthe providers you connect |
| Routing, retries and failover | yes | yes |
| Your own provider accounts | yesyou run it with your keys | yesconnected per environment |
| Self-hosting and open source | yesApache-2.0 | partialself-hosted by agreement; gateway not open source |
| Nothing to deploy or operate | nono hosted offering documented | yeshosted service |
| Budgets that refuse a request | yes402; virtual key, team, customer | yes402; org, project, member, key |
| Estimates at your own rates | yespricing overrides, nine scopes | yesstored per organization; Team plan |
| A model with no known price | partial$0 until an override is added | yesrecorded as unpriced; refused under a blocking budget by default |
| Reconciliation against the provider's own bill | no | partialOpenAI and Anthropic, Business plan; variance stated |
| Response caching | yessemantic | partialexact match, chat completions |
| MCP gateway | yesopen-source edition; per-key tool allow-lists | yespinned tools, grants per key; Business plan |
| Guardrails | partialEnterprise edition | partialchat patterns plus prompt classifiers; Business plan |
| Roles, SSO and audit log | partialEnterprise edition | yesaudit every plan; SSO and SCIM on Business |
| Image, video, and audio jobs and a media workspace | partialmedia request types; no workspace | yesasync jobs, signed webhooks, INFRO Studio |
| Available today | yes | noprivate early access |
Our reading of each product's public documentation as of October 2026. We do not quote competitors' prices — theirs change and a wrong number about someone else's business is worse than none. Think a row is wrong? Tell us and we will fix it.
Work out what the difference is worth
Estimate what your current model mix costs at your providers' list rates and the plan it needs, or have us read your real usage and tell you where we would and would not help.