Keep keys and data in your environment
- Store every provider credential inside the gateway, never in client code
- Keep all request data within your own infrastructure
- Rotate provider keys without touching a single application
Try Bifrost Enterprise free for 14 days.
Request access[ AI GATEWAY FOR COMPLIANCE OFFICERS ]Turn your AI policy into
Turn your AI policy into
an audit trail of every AI call
Bifrost logs each AI request you route through it with the user, team, and model behind it, so AI compliance evidence comes from one place.
Enterprise ready: VPC | On-prem | Air-gapped
Applications, agents, and MCP servers → Bifrost AI gateway → providers, with records to your database, bucket, SIEM, and GRC
Request
Virtual key
Cache
Route
Guardrail
Provider
Response
[ TRUSTED BY TEAMS BUILDING WITH BIFROST ]
[ WHY COMPLIANCE OFFICERS NEED AN AI GATEWAY ]
An AI acceptable use policy leaves no audit trail
Declared inventories and signed policies describe AI use, while the requests themselves are recorded in separate provider accounts.
The inventory depends on what teams report
When an AI register is built from intake forms and surveys, AI inside vendor products and models called directly from code reach the AI inventory only when someone declares them.
An acceptable use policy needs operating evidence
A policy states which data may reach which model. Testing whether that control operated takes a record of each request, its sender, and its model.
Retention requirements pull in opposite directions
FINRA notes that GenAI use can implicate broker-dealer recordkeeping rules, discovery exposure argues for short retention elsewhere, and prompts from personal accounts fall outside both policies.
Reviews start after the provider is chosen
When leadership selects an AI provider first, compliance has little time to review its data handling, approvals, and recordkeeping across the teams adopting it.
[ HOW IT WORKS ]
How an AI gateway supports AI governance compliance
An AI gateway sits between applications and model providers, so each request is checked and recorded at one point.
Model approval becomes a deny-by-default allowlist
When virtual keys are required on every request, each key lists the models its holder may call and denies the rest, under governance controls you set once.
Your identity provider names every sender
The virtual key, the signed-in user, and an optional project header identify who sent each request and why.
Audit logging into your log store and SIEM
Request metadata stays in your log store, payloads move to your bucket, and audit events export to your SIEM.
[ GATEWAY CAPABILITIES FOR COMPLIANCE OFFICERS ]
AI usage policy controls with audit logging built in
The gateway turns an AI usage policy into enforced rules and records a reviewer can test.
Log a correlation ID and sender on every AI request
- Record user ID, team, customer, business unit, and virtual key on each request
- Tag each request to a use case with a project, carried into every export
- Reject requests missing a required header such as X-Correlation-ID with HTTP 400
Approve models and providers before anyone can call them
- Deny every provider not listed on a virtual key, including providers added later
- Deny any model outside a listed provider's allowed_models list
- Expose no MCP tools to a virtual key until tools are configured, unless a client is marked Allow by Default
PII redaction and data masking in every AI record
- Write reversible placeholders to logs while live requests pass unchanged (logs_only)
- Detect PII with Azure AI Language, including its optional phi domain for health data
- Turn off content logging with disable_content_logging, gateway-wide or per virtual key
Set a data retention policy for AI records
- Set log_retention_days from 1 day upward; the default is 365
- Write prompts and responses to your S3 or GCS bucket under its lifecycle rules
- With content logging off, keep payloads in the bucket and out of the logs database
Configuration audit with S3 Object Lock
- Record who changed which resource, when, and whether the action succeeded
- Sign each audit event with an HMAC key so entries can be verified
- Archive audit events as JSON Lines to an S3-compatible bucket kept under S3 Object Lock
An auditor role with read-only access to AI records
- Create an Auditor role that can view audit logs, request logs, and guardrail settings
- Limit visible records to own, team, business unit, customer, or all data with data access control
- Remove a user when your identity provider deactivates or deletes them over SCIM
[ FITS YOUR EXISTING STACK ]
Bifrost works with your GRC tools, Purview, and archive
Bifrost feeds request records into the compliance systems you already run.
Your GRC platform stays the system of record
Vanta, OneTrust, or ServiceNow keeps your controls, with gateway exports attached as operating evidence.
Microsoft Purview keeps its Microsoft coverage
Purview keeps capturing Copilot and the enterprise AI apps it supports, while the gateway records API, agent, and MCP traffic from your applications.
Archives and SIEMs receive the same records
Audit events export in Syslog format for your SIEM, and a BigQuery export delivers one row per request to review beside Smarsh or Global Relay archives.
[ DEPLOYMENT ]
Keep every AI record in infrastructure you operate
Prompts, responses, and audit records stay in storage you run with Bifrost Enterprise.
Deploy anywhere, including air-gapped
- Run fully in your VPC or air-gapped private network
- Deploy with no external dependencies or outbound calls
- Self-host under Apache 2.0, or run fully managed
Meet enterprise compliance standards
- Comply with SOC 2 Type II, HIPAA, GDPR, and ISO 27001
- Maintain audit-ready records across every request
- Enforce one consistent policy across teams and providers
[ COMPLIANCE ]
Frameworks, attestations, and where gateway records fit
ISO/IEC 42001 control A.6.2.8 asks for AI event logs at least while systems are in use. Gateway records can support that evidence, and certification stays your auditor's decision.
Maxim AI, the company behind Bifrost, states it is SOC 2 Type II and ISO 27001 compliant. Request its reports through the trust center.
Company attestations
[ QUICK SETUP ]
Drop-in replacement for any AI SDK
Change one base URL. Bifrost is a drop-in replacement for the OpenAI, Anthropic, and Google GenAI SDKs, and works with LiteLLM, LangChain, and more.
shell
npx -y @maximhq/bifrostpython
client = anthropic.Anthropic(
base_url="http://localhost:8080/anthropic", # only change needed
api_key="<YOUR-BIFROST-VIRTUAL-KEY>",
)Source code is on GitHub under Apache 2.0, built by Maxim AI.
[ FAQ ]
Frequently asked questions
AI governance compliance is showing, with records, that an organization's AI use follows its policies, approvals, and applicable regulations. Bifrost supports it at the request level: models are denied until approved on a virtual key, and each allowed request records its sender, model, and use case through a project.
ISO/IEC 42001 is the international standard for establishing, running, and improving an AI management system. Its Annex A control A.6.2.8 asks for AI event logs at least while a system is in use, and Bifrost request logs can supply them for traffic routed through the gateway.
Under the Digital Omnibus on AI, Regulation (EU) 2026/1744, the EU AI Act's obligations for stand-alone high-risk systems apply from December 2, 2027, and for high-risk AI embedded in regulated products from August 2, 2028. Deployers keep high-risk system logs they control for at least six months under Article 26.
An AI acceptable use policy is enforced by applying its rules to each request before it reaches a model. In Bifrost, approved models are listed per virtual key, MCP tools stay denied by default until configured, required headers such as a correlation ID return HTTP 400 when missing, and guardrails redact personal data, so an out-of-policy request is refused before it reaches a model.
Logged prompts are electronically stored information, which Federal Rule of Civil Procedure 34 lets a party request in US litigation, while FINRA notes that GenAI use can implicate recordkeeping rules. Storage should follow your retention schedule. Bifrost separates metadata from content: content logging can be turned off gateway-wide, per virtual key, or per connector, and request-log retention is set in days, applied gateway-wide rather than per use case. See log exports.
Access to prompt logs should be limited to named roles with a review purpose. Bifrost Enterprise applies role-based permissions to request, audit, and MCP logs, data access scopes for own, team, business unit, customer, or all data, and a separate Reveal permission for redacted values. Users your identity provider deactivates over SCIM are removed.
An AI gateway records the model and tool calls your own applications and agents send. Purview keeps covering Copilot and its supported enterprise AI apps, the GRC platform holds control documentation, and Bifrost exports request records (for example, through BigQuery) and audit events as evidence.
Bifrost audit logs record administrative changes and can be signed with an HMAC key so entries can be verified. The docs recommend S3 Object Lock (WORM) and bucket lifecycle rules to govern how long the audit archive is kept. Request logs follow your retention settings, so long-term request records also belong in a locked bucket.
[ RELATED RESOURCES ]
Related resources
Where the controls behind these records are documented.
Test the gateway against a recent audit request
To test Bifrost for AI governance compliance, route one regulated workflow through the gateway and check its records against a recent audit request. That pilot fits within the 14-day free Enterprise trial.
[ BIFROST FEATURES ]
Open Source & Enterprise
Everything you need to run AI in production, from free open source to enterprise-grade features.
01 Governance
SAML support for SSO and Role-based access control and policy enforcement for team collaboration.
02 Adaptive Load Balancing
Automatically optimizes traffic distribution across provider keys and models based on real-time performance metrics.
03 Cluster Mode
High availability deployment with automatic failover and load balancing. Peer-to-peer clustering where every instance is equal.
04 Alerts
Real-time notifications for budget limits, failures, and performance issues on Email, Slack, PagerDuty, Teams, Webhook and more.
05 Log Exports
Export and analyze request logs, traces, and telemetry data from Bifrost with enterprise-grade data export capabilities for compliance, monitoring, and analytics.
06 Audit Logs
Comprehensive logging and audit trails for compliance and debugging.
07 Vault Support
Secure API key management with HashiCorp Vault, AWS Secrets Manager, Google Secret Manager, and Azure Key Vault integration.
08 VPC Deployment
Deploy Bifrost within your private cloud infrastructure with VPC isolation, custom networking, and enhanced security controls.
09 Guardrails
Automatically detect and block unsafe model outputs with real-time policy enforcement and content moderation across all agents.
[ SHIP RELIABLE AI ]
Try Bifrost Enterprise with a 14-day Free Trial
[quick setup]
Drop-in replacement for any AI SDK
Change just one line of code. Works with OpenAI, Anthropic, Vercel AI SDK, LangChain, and more.
1import os
2from anthropic import Anthropic
3
4anthropic = Anthropic(
5 api_key=os.environ.get("ANTHROPIC_API_KEY"),
6 base_url="https://<bifrost_url>/anthropic",
7)
8
9message = anthropic.messages.create(
10 model="claude-3-5-sonnet-20241022",
11 max_tokens=1024,
12 messages=[
13 {"role": "user", "content": "Hello, Claude"}
14 ]
15)
Drop in once, run everywhere.