AI Gateway vs. Bifrost (Maxim AI): Run a Go binary yourself or a governance component in your own tenant?

Key facts (as of 26 September 2026):

• Bifrost by Maxim AI is, per its documentation, a self-operated AI gateway written in Go (binary, Docker image or Helm chart) whose core is licensed under Apache License 2.0, while SSO, RBAC, guardrails, audit logs, clustering and in-VPC support are assigned by the vendor to a separately licensed Enterprise tier.
• The innFactory AI Gateway is deployed into the customer’s cloud tenant and bundles cost centers with per-member caps, budgets in euros and a companion that measures the agent time of Claude Code, Codex and OpenCode per repository, including subscription usage.
• Per its documentation, Bifrost keeps budgets per customer, team, virtual key and provider configuration and logs agent requests server-side in the gateway log; a local companion recording agent time independently of the gateway log is not described in Bifrost’s documentation (as of 26 September 2026), whereas the innFactory AI Gateway includes SSO, seven roles and guardrails in every license tier and ships the companion.
• Bifrost fits when you already run Kubernetes, want to own a gateway rolled out via Helm and need SAML, SCIM, air-gapped operation or external guardrail providers; the innFactory AI Gateway fits when cost centers with per-member caps, EUR at the ECB rate, agent time including subscription usage and the Flow Builder come first and you do not want to build the rollout yourself.

Bifrost is an open-source AI gateway by Maxim AI. Its documentation describes it as “A high-performance AI gateway unifying 20+ providers through a single OpenAI-compatible API” - a proxy written in Go that you run yourself. Run as a binary, Docker image or Helm chart and extended with an Enterprise tier when needed, it is a solid, well-documented product.

The innFactory AI Gateway starts from a different place: not a library for self-operation but an infrastructure component that innFactory rolls out into your cloud tenant - with a FinOps layer of cost centers, per-member caps and EUR display, and a companion that records the working time of coding agents even when they run on a subscription and never touch the gateway.

So the decision is not “which gateway is better” but: operate an open-source core with Enterprise extensions yourself - or a governance and FinOps component in your own tenant that shares its stack with CompanyGPT?

What Bifrost is

Bifrost is developed by Maxim AI; according to the privacy policy, the legal entity is H3 Labs Inc, based in Mountain View, California. The core is licensed under Apache License 2.0 (“Copyright 2025 H3 Labs Inc.”). An acquisition is not documented.

The pricing page lists routing, fallbacks, virtual keys, budgets, the MCP gateway, semantic caching and observability, among others, as open source. Alongside it there is a separately licensed Enterprise tier: “Enterprise features such as adaptive load balancing, clustering, guardrails, RBAC, SSO, audit logs, and in-VPC support are licensed separately.” A large part of what companies understand as governance therefore sits in Bifrost’s Enterprise tier.

Operating model: SaaS, self-hosting or infrastructure in your own tenant

You run Bifrost yourself: per the README via npx -y @maximhq/bifrost or docker run maximhq/bifrost, plus a Go SDK, and on Kubernetes via a documented Helm chart with example values up to production HA. The Enterprise tier adds a cluster mode (“Peer-to-peer clustering where every instance is equal”) plus in-VPC, air-gapped and multi-cloud operation. We found no vendor-hosted SaaS gateway; the vendor stresses: “Keep all request data within your own infrastructure.”

The innFactory AI Gateway also runs on your side, but innFactory does not host it: it rolls the gateway out as a cloud-native component into your tenant - Azure Container Apps or AKS as the reference path, alternatively STACKIT or any Kubernetes. The backend is a Rust binary in a container plus a dashboard image, backed by PostgreSQL (for example STACKIT PostgreSQL Flex), MongoDB or Azure Cosmos DB; the data path stays in the tenant, the configuration with you as YAML or Terraform. innFactory GmbH, a Microsoft CSP, supports the rollout and can provide the Azure subscription. We currently ship no Helm chart; here Bifrost is more immediate for Kubernetes teams.

Identity, SSO and roles

According to its documentation, Bifrost connects identity providers via OIDC - Okta, Microsoft Entra, Keycloak, Zitadel, Auth0, Google Workspace “or any OIDC provider”, using Authorization Code plus PKCE. Add SAML and inbound SCIM 2.0 via a /scim/v2 API. Roles: Admin, Developer, Viewer and custom roles. The pricing page assigns all of this to the Enterprise tier; a user limit for SSO in the open-source scope is not documented. Models, budgets and limits are assigned through virtual keys “per user, team, or business unit”.

In the innFactory AI Gateway, identity integration is part of every license tier: Microsoft Entra ID or generic OIDC (Keycloak documented, AWS Cognito claims handled), OIDC/JWT on the proxy path optional. Seven roles - Owner, Admin, Auditor, Finance, Security, Member, Guest - with IdP group mapping. Finance sees costs and margins, Security sees guardrails and audit log; cost centers have their own administrators, and logins from foreign e-mail domains automatically become Guests without proxy access. We do not include SAML and SCIM; if you strictly need them, Bifrost’s Enterprise tier has the broader catalog.

Cost control and AI FinOps

According to its documentation, Bifrost organizes budgets in a hierarchy of customer, team, virtual key and provider configuration; each level has its own budget. A request only runs if every affected level still has balance, after which the cost is deducted from all levels. Reset windows range from one minute to one year, calendar-aligned to UTC. Basic budgets with virtual keys are open source; hierarchical budgets, alerts (“on Email, Slack, PagerDuty, Teams, Webhook and more”), log exports and audit logs are Enterprise. Not documented are currency, per-request cost reporting and the source of the model prices.

The innFactory AI Gateway is deliberately deeper here. Four independent gates check every request: organization (monthly), cost center (monthly plus daily and weekly windows, pool or per-user mode with caps per member), user and API key. When exhausted, each gate returns HTTP 402 with error code E4006 or optionally routes to a budget fallback model; per-member caps are always hard stops. Add per-model caps, TPM and RPM limits, alerts at 80 and 90 percent and a budget_exceeded event for flows and webhooks. Costs are attributed by user, key, team, cost center, model, provider, tool, repository and billing source (gateway or subscription), kept in USD microdollars and shown in EUR at the daily ECB rate; model prices come from ai-prices.eu after admin review. Every buffered response carries x-aigateway-response-cost; Insights and CSV export include agent and LLM time.

Coding agents, subscriptions and agent time

Bifrost documents integrations for Claude Code, Codex CLI, Gemini CLI, Qwen Code, Opencode, GitHub Copilot, Cursor, Zed, Roo Code, LibreChat, Open WebUI, Claude for Office and Claude Desktop. A dedicated Bifrost CLI connects agents “with one command”; measurement is server-side: “All agent interactions are automatically logged and can be monitored at http://localhost:8080/logs.” According to Bifrost’s documentation (as of 26 September 2026), a local companion that records agent working time independently of the gateway log is not described; nor is an MDM rollout of a client component.

The innFactory AI Gateway offers six proxy setups: Claude Code, Codex, OpenCode, Cowork (macOS .mobileconfig, Windows .reg), VS Code Continue and VS Code Copilot (BYOK). Admins create templates that each developer clones with their own key and cost center; aigw run claude|codex|opencode launches the tool with a short-lived proxy token.

The real difference is the companion aigw: a small background service for Linux, macOS and Windows that reads the local session logs of Claude Code, Codex and OpenCode and reports agent time, LLM time and token classes per repository, user, tool and cost center. It detects whether a session ran through the gateway or on the tool’s subscription (Claude Max, Codex plan); subscription usage is valued as shadow tokens at list price but never billed, budgeted or rate-limited, and always shown separately. Rollout runs via device-code login or enrollment tokens through Intune, macOS policy or a Linux script; the upload is a code-frozen allow-list without prompts, absolute paths or remote URLs with credentials. The companion does not cover other tools such as Cursor or Gemini CLI, and the binaries are not yet code-signed.

Providers, dialects and models

Bifrost is broader here. The documentation names more than 20 providers and more than 1,000 models, including Anthropic, Azure, Bedrock, Cohere, DeepSeek, ElevenLabs, Gemini, Groq, Mistral, Ollama, OpenAI, OpenRouter, Runway, Vertex AI, vLLM and xAI. STACKIT and IONOS are not listed in the provider reference (as of 26 September 2026). The primary interface is the OpenAI-compatible API; according to the documentation there is also an Anthropic-Messages-compatible mode and the Responses API. Bifrost positions itself as a drop-in for the OpenAI, Anthropic, Bedrock and Google GenAI SDKs.

The innFactory AI Gateway knows 13 provider types: OpenAI, Azure OpenAI, Azure AI Foundry, Anthropic (direct or Vertex), Gemini (AI Studio or Vertex including the EU endpoint), Mistral, AWS Bedrock including Mantle, STACKIT, IONOS, Ollama, LM Studio, ElevenLabs and OpenAI-compatible endpoints. Three inbound dialects - OpenAI, Anthropic, Gemini - plus realtime WebSocket including Gemini Live via Vertex EU, audio, embeddings, images, OCR, and files and batches. If you need many specialist providers, Bifrost serves you better; if you want STACKIT or IONOS as sovereign providers, you find them with us.

Governance: guardrails, MCP, routing and Flow Builder

Guardrails. Enterprise-tier in Bifrost: Presidio-based PII detection and redaction, prompt injection detection, jailbreak protection via Azure Content Safety, and connections to AWS Bedrock Guardrails, Google Model Armor and Patronus AI. The redaction modes runtime, logs_only and runtime_reversible are clearly described; a named shadow mode is not documented. The vendor describes the audit log as aligned with frameworks such as SOC 2, GDPR, HIPAA and ISO 27001.

The innFactory AI Gateway ships nine rule types in every license - secret detection, PII block or redact, banned keywords, token limit including maximum cost, Azure Content Safety, prompt injection, denied topics, groundedness, external webhook - with the actions block, redact, warn and audit, a shadow mode, mid-stream redaction, a GDPR preset and a /check dry run.

MCP. Bifrost has a native MCP gateway: servers are registered once, the gateway authenticates via OAuth 2.0 at server or user level and scopes tool access via virtual keys and SSO roles; the basic gateway is open source, “MCP with Federated Auth” is Enterprise. Our MCP proxy supports eight upstream auth modes - from static headers via per-user PKCE and Entra on-behalf-of to AWS SigV4 - plus toolsets, semantic tool search, guardrails on tool arguments and audit per call.

Routing and flows. According to its documentation, Bifrost distributes requests weighted per virtual key, creates fallback chains automatically and classifies prompts into three tiers via a complexity router; Enterprise adds adaptive load balancing. Configuration is via web UI, API or config.json; a visual flow builder is not documented. The innFactory AI Gateway has an auto-router, aliases, fallbacks and deployment groups - and on top the Flow Builder: a visual graph with nine strategies (including least_cost, cost_cascade, capability_match), guardrail nodes with pass/block/fail branching and event triggers that fire webhooks. A flow is callable like a model and bills the same budgets.

What Bifrost does better

• Provider breadth: more than 20 documented providers including image, video, speech and inference platforms such as vLLM.
• Kubernetes: Helm chart with example values and an Enterprise cluster mode.
• Deployment: in-VPC, air-gapped and multi-cloud operation.
• Identity: SAML in addition to OIDC, inbound SCIM 2.0.
• External guardrail providers: AWS Bedrock Guardrails, Google Model Armor, Patronus AI.
• Secret managers: HashiCorp Vault, AWS, Google and Azure Key Vault in the Enterprise tier.
• Agent ecosystem: guides for 13 agents and interfaces plus a dedicated CLI.
• Support: SLA-backed Enterprise support according to the pricing page.

When Bifrost is the right choice

• You already run Kubernetes and want to roll out a gateway via Helm that your team owns.
• You need SAML or SCIM and license the Enterprise tier for it.
• You connect many specialist providers - inference platforms, image and video models.
• You already use guardrail services from AWS, Google or Patronus.
• You need air-gapped operation.

When the innFactory AI Gateway is the right choice

• You want to know how much working time coding agents deliver - including on Claude Max or a Codex plan.
• You need cost centers with caps per member, EUR at the ECB rate and cost per request.
• You want SSO, seven roles and guardrails in every license tier.
• You want to model routing, guardrails and webhooks visually in the Flow Builder.
• You use CompanyGPT and want one shared stack of identity, cost centers, guardrails and MCP registry.
• You want the gateway in your own tenant but do not want to build the rollout yourself.

Comparison at a glance

Conclusion

Bifrost is a good product for teams that want an AI gateway as an open-source building block in their own Kubernetes or VPC landscape, adding governance through the Enterprise tier when needed. Provider breadth, Helm chart, SAML, SCIM and connections to AWS Bedrock Guardrails, Google Model Armor and Patronus AI Bifrost has; we currently do not.

The innFactory AI Gateway is the right choice when the question is not “which proxy” but “who pays for what, and how much agent work really happens”. Cost centers with member caps, EUR at the ECB rate, the companion for agent time including subscription usage and the Flow Builder answer exactly that - as a component in your tenant, rolled out by innFactory.

Natively integrated with CompanyGPT, one reusable stack

The AI Gateway is natively integrated with CompanyGPT: the same budgets, cost centers and guardrails apply to chat, agents and add-ons. Identity via Entra ID or Keycloak, cost centers, guardrails and the MCP registry form one stack that is reused across CompanyGPT, the gateway and your own tools. The gateway also runs standalone. See all comparisons and the AI Gateway product page.

Sources

Retrieved on 26 September 2026:

• Bifrost product page — https://getmaxim.ai/bifrost
• Bifrost repository and README — https://github.com/maximhq/bifrost
• Bifrost LICENSE (Apache 2.0) — https://github.com/maximhq/bifrost/blob/main/LICENSE
• Bifrost documentation, home — https://docs.getbifrost.ai
• Bifrost documentation, overview — https://docs.getbifrost.ai/overview
• Bifrost pricing (tier scope) — https://www.getmaxim.ai/bifrost/pricing
• Bifrost Enterprise — https://www.getmaxim.ai/bifrost/enterprise
• Enterprise deployment — https://www.getmaxim.ai/bifrost/resources/enterprise-deployment
• Helm deployment guide — https://docs.getbifrost.ai/deployment-guides/helm
• Enterprise governance (SSO, SCIM, roles) — https://docs.getbifrost.ai/enterprise/governance
• Budgets and limits — https://docs.getbifrost.ai/features/governance/budget-and-limits
• Routing — https://docs.getbifrost.ai/features/governance/routing
• Guardrails — https://docs.getbifrost.ai/enterprise/guardrails
• MCP gateway — https://www.getmaxim.ai/mcp-gateway
• CLI agents overview — https://docs.getbifrost.ai/cli-agents/overview
• Supported providers — https://docs.getbifrost.ai/providers/supported-providers
• Anthropic SDK integration — https://docs.getbifrost.ai/integrations/anthropic-sdk/overview
• Changelog v2.0.0 (Responses API) — https://docs.getbifrost.ai/changelogs/v2.0.0
• Maxim AI privacy policy (legal entity) — https://getmaxim.ai/privacy-policy

Note on the information: All statements about other vendors’ products are based on their publicly available documentation as of the stated date. Vendors continuously develop their products, features and terms — the vendor’s current information always prevails. If any statement appears outdated or incorrect to you, please drop us a line at info@innfactory.ai; we will review and correct it promptly. This comparison does not replace legal or data-protection advice in individual cases.

Further reading

• AI Gateway – proxy, FinOps layer and companion at a glance
• CompanyGPT – the chat platform that shares the stack with the gateway
• Claude Code as an AI Harness – the coding agent the companion measures
• AI Harness explained – fundamentals of agent harnesses
• All AI Gateway comparisons – LiteLLM, OpenRouter, Portkey, Cloudflare, Kong, Bifrost, TrueFoundry, Helicone, Vercel
• AI Gateway vs. LiteLLM
• AI Gateway vs. Helicone
• AI Gateway vs. Kong AI Gateway