Client Configuration

Configure the Bifrost client: connection pool, logging, CORS, header filtering, compat shims, and MCP settings

The bifrost.client block controls how Bifrost manages its internal worker pool, request logging, authentication enforcement, header policies, SDK compatibility shims, and MCP agent behaviour. All settings map directly to the client section of the rendered config.json.

---

Connection Pool

Parameter	Description	Default
bifrost.client.initialPoolSize	Pre-allocated worker goroutines per provider queue	300
bifrost.client.dropExcessRequests	Drop requests when queue is full instead of waiting	false

A larger pool reduces latency spikes under burst load at the cost of higher baseline memory. For production workloads with multiple providers, 1000 is a common starting point.

# client-pool.yaml
image:
  tag: "v1.4.11"

bifrost:
  client:
    initialPoolSize: 1000
    dropExcessRequests: true   # Return 429 instead of queuing indefinitely

helm install bifrost bifrost/bifrost -f client-pool.yaml

# Or set inline
helm upgrade bifrost bifrost/bifrost \
  --reuse-values \
  --set bifrost.client.initialPoolSize=1000 \
  --set bifrost.client.dropExcessRequests=true

---

Request & Response Logging

Parameter	Description	Default
bifrost.client.enableLogging	Log all LLM requests and responses	true
bifrost.client.disableContentLogging	Strip message content from logs (keeps metadata)	false
bifrost.client.logRetentionDays	Days to retain log entries in the store	365
bifrost.client.loggingHeaders	HTTP request headers to capture in log metadata	[]

Set disableContentLogging: true for HIPAA / PCI compliance workloads where message content must not be persisted.

bifrost:
  client:
    enableLogging: true
    disableContentLogging: true    # PII / compliance: store metadata only
    logRetentionDays: 90
    loggingHeaders:
      - "x-request-id"
      - "x-user-id"

helm upgrade bifrost bifrost/bifrost \
  --reuse-values \
  --set bifrost.client.disableContentLogging=true \
  --set bifrost.client.logRetentionDays=90

---

Security & CORS

Parameter	Description	Default
bifrost.client.allowedOrigins	CORS allowed origins	["*"]
bifrost.client.enforceGovernanceHeader	Require x-bf-vk virtual-key header on every request	false
bifrost.client.maxRequestBodySizeMb	Maximum allowed request body size	100
bifrost.client.whitelistedRoutes	Routes that bypass auth middleware	[]

bifrost:
  client:
    allowedOrigins:
      - "https://app.yourdomain.com"
      - "https://admin.yourdomain.com"
    enforceGovernanceHeader: true  # Every request must carry a virtual key
    maxRequestBodySizeMb: 50
    whitelistedRoutes:
      - "/health"
      - "/metrics"

helm install bifrost bifrost/bifrost \
  --set image.tag=v1.4.11 \
  --set bifrost.client.enforceGovernanceHeader=true

---

Header Filtering

Controls which x-bf-eh-* headers are forwarded to upstream LLM providers.

Parameter	Description	Default
bifrost.client.headerFilterConfig.allowlist	Only these headers are forwarded (whitelist mode)	[]
bifrost.client.headerFilterConfig.denylist	These headers are always blocked	[]
bifrost.client.requiredHeaders	Headers that must be present on every request	[]
bifrost.client.allowedHeaders	Additional headers permitted for CORS and WebSocket	[]

When both lists are empty, all x-bf-eh-* headers pass through. Specifying an allowlist enables strict whitelist mode - only listed headers are forwarded.

bifrost:
  client:
    headerFilterConfig:
      allowlist:
        - "x-bf-eh-anthropic-version"
        - "x-bf-eh-openai-beta"
      denylist: []
    requiredHeaders:
      - "x-request-id"

---

Authentication

Parameter	Description	Default
bifrost.authConfig.isEnabled	Enable username/password auth for the API and dashboard	false
bifrost.authConfig.adminUsername	Admin username (plain text, prefer secret)	""
bifrost.authConfig.adminPassword	Admin password (plain text, prefer secret)	""
bifrost.authConfig.existingSecret	Kubernetes Secret name for credentials	""
bifrost.authConfig.usernameKey	Key within the secret for username	"username"
bifrost.authConfig.passwordKey	Key within the secret for password	"password"
bifrost.authConfig.disableAuthOnInference	Skip auth check on /v1/* inference routes	false

# Create secret first
kubectl create secret generic bifrost-admin \
  --from-literal=username='admin' \
  --from-literal=password='your-secure-password'

bifrost:
  authConfig:
    isEnabled: true
    disableAuthOnInference: false
    existingSecret: "bifrost-admin"
    usernameKey: "username"
    passwordKey: "password"

helm upgrade bifrost bifrost/bifrost \
  --reuse-values \
  -f auth-values.yaml

---

Encryption

Parameter	Description	Default
bifrost.encryptionKey	Optional encryption key (plain text - use encryptionKeySecret in production). If omitted, data is stored in plaintext.	""
bifrost.encryptionKeySecret.name	Kubernetes Secret name containing the key	""
bifrost.encryptionKeySecret.key	Key within the secret	"encryption-key"

Always use a Kubernetes Secret in production:

kubectl create secret generic bifrost-encryption \
  --from-literal=encryption-key='your-32-byte-encryption-key-here'

bifrost:
  encryptionKeySecret:
    name: "bifrost-encryption"
    key: "encryption-key"

helm install bifrost bifrost/bifrost \
  --set image.tag=v1.4.11 \
  -f encryption-values.yaml

---

Async Jobs & Database Pings

Parameter	Description	Default
bifrost.client.disableDbPingsInHealth	Exclude DB connectivity from /health checks	false
bifrost.client.asyncJobResultTTL	TTL (seconds) for async job results	3600

---

Compat Shims

Compatibility flags that let Bifrost silently adapt request/response shapes for SDK integrations:

Parameter	Description	Default
bifrost.client.compat.convertTextToChat	Wrap legacy text completions as chat messages	false
bifrost.client.compat.convertChatToResponses	Translate chat completions to Responses API format	false
bifrost.client.compat.shouldDropParams	Silently drop unsupported parameters instead of erroring	false
bifrost.client.compat.shouldConvertParams	Auto-convert parameter names across provider schemas	false
bifrost.client.compat.forceReasoningOnlyModelsToResponses	Route chat completions to the Responses API for models that cannot reason with tool calls, so reasoning is kept instead of forced off	true

bifrost:
  client:
    compat:
      shouldDropParams: true     # Useful when proxying mixed SDK traffic
      convertTextToChat: true    # For clients using the legacy /v1/completions endpoint

---

Prometheus Labels

Add custom labels to every Prometheus metric emitted by Bifrost:

bifrost:
  client:
    prometheusLabels:
      - name: "environment"
        value: "production"
      - name: "region"
        value: "us-east-1"

---

MCP Agent Settings

Parameter	Description	Default
bifrost.mcp.toolManagerConfig.maxAgentDepth	Maximum tool-call recursion depth for MCP agent mode	10
bifrost.mcp.toolManagerConfig.toolExecutionTimeout	Timeout per tool execution in seconds	30
bifrost.mcp.toolManagerConfig.codeModeBindingLevel	Code mode binding level (server or tool)	server
bifrost.mcp.toolManagerConfig.disableAutoToolInject	Disable automatic MCP tool injection	false
bifrost.mcp.toolSyncInterval	Global tool sync interval as a Go duration string in whole minutes (for example 10m). Use 0s to use the runtime default of 10 minutes; periodic sync cannot be turned off. Legacy bifrost.client.mcpToolSyncInterval: 0 means the same default.	10m

bifrost:
  mcp:
    toolSyncInterval: "15m"
    toolManagerConfig:
      maxAgentDepth: 15
      toolExecutionTimeout: 60
      codeModeBindingLevel: "tool"
      disableAutoToolInject: false

---

Full Example

# client-full.yaml
image:
  tag: "v1.4.11"

bifrost:
  encryptionKeySecret:
    name: "bifrost-encryption"
    key: "encryption-key"

  authConfig:
    isEnabled: true
    disableAuthOnInference: false
    existingSecret: "bifrost-admin"
    usernameKey: "username"
    passwordKey: "password"

  client:
    initialPoolSize: 1000
    dropExcessRequests: true
    allowedOrigins:
      - "https://app.yourdomain.com"
    enableLogging: true
    disableContentLogging: false
    logRetentionDays: 90
    enforceGovernanceHeader: true
    maxRequestBodySizeMb: 100
    headerFilterConfig:
      allowlist: []
      denylist: []
    prometheusLabels:
      - name: "environment"
        value: "production"
  mcp:
    toolSyncInterval: "10m"
    toolManagerConfig:
      maxAgentDepth: 10
      toolExecutionTimeout: 30
      codeModeBindingLevel: "server"
      disableAutoToolInject: false

# Create prerequisites
kubectl create secret generic bifrost-encryption \
  --from-literal=encryption-key='your-32-byte-encryption-key-here'

kubectl create secret generic bifrost-admin \
  --from-literal=username='admin' \
  --from-literal=password='your-secure-password'

# Install
helm install bifrost bifrost/bifrost -f client-full.yaml

Agent Gateway destination permissions

Helm value	Rendered config.json field	Default
bifrost.client.a2aPrivateAllowlist	client.a2a_private_allowlist	[]
bifrost.agents[].credential_origins	agents[].credential_origins	[]

The global allowlist permits private or loopback hosts, IP addresses, and CIDR ranges for card discovery, OAuth, and runtime HTTP or gRPC connections. Entries must not contain a scheme, port, or path. Link-local and cloud metadata destinations remain blocked.

The per-agent list permits runtime origins outside the card’s own origin, which is automatically permitted. Use an HTTP(S) scheme, hostname, and optional port without a path, userinfo, query, or fragment. Up to 32 unique additional origins are accepted. This check applies even without configured runtime authentication.

For a public card that advertises a private runtime at https://research.internal:9000, configure both:

bifrost:
  client:
    a2aPrivateAllowlist:
      - research.internal
  agents:
    - name: library-research
      agent_card_url: https://cards.example.com/.well-known/agent-card.json
      credential_origins:
        - https://research.internal:9000

Adding an origin does not grant private-network access, and adding a host to the network allowlist does not approve a different runtime origin. If the card or OAuth server is also private, include its destination in a2aPrivateAllowlist too. See Agent Gateway authentication for credentials and TLS examples.

Agent entries use the config.json field names unchanged, including credential_origins; the client value uses the chart’s camelCase name. Registrations are reconciled at startup. See source-of-truth behavior before managing the same agent through Helm and the dashboard.