Changelog
- Added
serviceMonitorto create a Prometheus Operator ServiceMonitor that scrapes Bifrost’s/metrics. Off by default. - Added
serviceMonitor.basicAuthandserviceMonitor.authorizationfor scraping when auth is enabled.basicAuthreuses your existing auth Secret. - The chart now refuses
serviceMonitor.enabled: truewhen telemetry metrics are turned off. - The main Service now carries the
app.kubernetes.io/component: serverlabel, so the ServiceMonitor scrapes only that Service. - Added
bifrost.mcp.clientConfigs[].serverTypeOnSlugto turn code mode on or off for a client on its own/mcp/<slug>endpoint. - Added
bifrost.client.mcpClientMetadataUrlto host Bifrost’s MCP Client ID Metadata Document elsewhere, for MCP servers that cannot reach Bifrost. - Added
bifrost.modelRoutersto configure model routers, which classify requests as simple, medium or complex and send each tier to a chosen model. They run after routing rules. With config.json as the source of truth,[]removes stored routers. bifrost.governance.complexityAnalyzerConfig.classifiernow acceptsllm, which requires thellmblock.- Added Prometheus metrics for the embedded Redis:
vectorStore.redis.metrics.enabledruns aredis_exportersidecar on port 9121. - Added
vectorStore.redis.metrics.serviceMonitorto create a Prometheus Operator ServiceMonitor for the Redis exporter, with optional basic auth. - Added
vectorStore.redis.master.extraEnvandenvFromto pass environment variables such asREDIS_ARGSto the embedded Redis. They work with Redis auth enabled. - Fixed the embedded Redis not saving its data to the persistent volume when Redis auth is enabled, which lost data whenever the pod was replaced. Behavior change: the Redis pod restarts once on upgrade.
- Added
bifrost.client.usageEstimationTokenizersto choose the tokenizers that estimate usage for cancelled or timed-out streams. - Added
bifrost.datasheetOverridesfor pricing and model-parameter overrides that survive datasheet syncs. With config.json as the source of truth,[]removes stored overrides. - Baseten keys now accept a per-alias
deployment_url. - The semantic cache now accepts
byteplusandbasetenas providers. - Added
bifrost.client.mcpOpenapiMaxSpecMb(client.mcp_openapi_max_spec_mb) to raise or lower the size limit on an openapi MCP client’s spec. 0 keeps the 32 MiB default. - Added
bifrost.client.bifrostUrl(client.bifrost_url), the gateway’s public base URL includingserver.basePath./v1/mcpneeds it; it accepts env and vault references. - Added
bifrost.client.bifrostMcpAllowWrites(client.bifrost_mcp_allow_writes) to stop/v1/mcpserving configuration changes as tools. On by default. - Added
bifrost.framework.pricing.openapiUrl(framework.pricing.openapi_url) to point/v1/mcpat a local OpenAPI spec on air-gapped installs.

